CN115762724A - Method, device and system for automatically delineating target area of medical image - Google Patents
Method, device and system for automatically delineating target area of medical image Download PDFInfo
- Publication number
- CN115762724A CN115762724A CN202211282046.8A CN202211282046A CN115762724A CN 115762724 A CN115762724 A CN 115762724A CN 202211282046 A CN202211282046 A CN 202211282046A CN 115762724 A CN115762724 A CN 115762724A
- Authority
- CN
- China
- Prior art keywords
- target area
- medical image
- target
- model
- sample set
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Apparatus For Radiation Diagnosis (AREA)
Abstract
The invention provides a method, a device and a system for automatically delineating a target area of a medical image. Firstly, positioning a target area of a medical image to be sketched by adopting a trained target area positioning model so as to acquire multi-channel information of the target area; and then, according to the multi-channel information, segmenting the medical image to be delineated by adopting a trained target segmentation model to obtain a target delineation result of the medical image to be delineated. The invention can realize the end-to-end flow, and has high sketching efficiency and strong universality; and the whole sketching process does not need manual participation, so that the problem of differentiation possibly generated by human factors can be reduced, and doctors can be better assisted to improve the quality and efficiency of radiotherapy plan design.
Description
Technical Field
The invention relates to the technical field of medical image processing, in particular to a method, a device and a system for automatically delineating a target area of a medical image.
Background
Tumor radiotherapy is a local treatment method for treating tumors by using radiation, about 70% of cancer patients need to be treated by using radiation therapy in the process of treating cancers, and about 40% of cancers can be radically treated by using radiation therapy. The role and position of radiotherapy in tumor therapy are increasingly highlighted, and the radiotherapy has become one of the main means for treating malignant tumors.
Delineation of the Target Volume (CTV) is the core work of the radiotherapy physician and plays an important role in controlling the tumor. A radiotherapy doctor delineates a treatment target area layer by layer based on a tomographic image, and then a physicist performs radiotherapy plan design according to the three-dimensional form of the target area and the relation with surrounding tissues, accurately calculates the irradiation dose, and concentrates the dose in the target area to the maximum extent so as to achieve the purposes of accurately irradiating a target area and protecting surrounding normal tissues and organs.
However, the standard CT positioning images usually contain dozens or even hundreds of slices, and require the doctor to manually delineate the treatment region in each slice individually, and the manual delineation method has the following defects:
1. because manual drawing is time-consuming and labor-consuming, a large amount of human resources are consumed in the whole drawing process, large-scale data is difficult to process through manual drawing, and the drawing efficiency is low;
2. the delineation quality of the target area depends on the experience and subjective judgment of a radiotherapy doctor, and the requirement on the radiotherapy doctor is high;
3. although the target area is drawn with uniform drawing consensus, different doctors have different understandings of the consensus, so that different doctors and even the same doctor may give different drawings to the same target area at different times, that is, the defect of manual drawing is the heterogeneity of the target areas drawn by different doctors. It has been found that when the target areas of different doctors are overlapped, the target areas are significantly different, which directly affects the dose distribution of the radiotherapy prescription and adversely affects the design of the radiotherapy plan.
It is noted that the information disclosed in this background of the invention section is only for enhancement of understanding of the general background of the invention and should not be taken as an acknowledgement or any form of suggestion that this information forms the prior art already known to a person skilled in the art.
Disclosure of Invention
The invention aims to provide a method, a device and a system for automatically delineating a target area of a medical image aiming at the defects of low target area delineating efficiency and poor consistency in the prior art.
In order to achieve the above object, the present invention is implemented by the following technical solution, a method for automatically delineating a target region of a medical image, comprising:
positioning a target area of a medical image to be delineated by adopting a trained target area positioning model to acquire multi-channel information of the target area, wherein the multi-channel information comprises morphological information and layer information corresponding to the morphological information;
and according to the multi-channel information, segmenting the medical image to be sketched by adopting a trained target region segmentation model to obtain a target region sketching result of the medical image to be sketched.
Optionally, the target area to be delineated medical image is positioned by using the trained target area positioning model to acquire the multi-channel information of the target area, including:
acquiring the morphological distribution of the target area by adopting the target area positioning model according to the slice image of the medical image to be sketched;
obtaining multi-channel information of the target area according to the morphological distribution of the target area; the multichannel information comprises a plurality of channel information, and each channel information comprises morphological information and layer information corresponding to the morphological information.
Optionally, the method for automatically delineating the target region of the medical image further comprises: and carrying out post-processing on the initial target area image obtained by segmenting the medical image to be sketched, and taking the post-processed initial target area image as the sketching result of the target area.
Optionally, the target area localization model is a neural network model; the method comprises the following steps of before positioning a target area of a medical image to be delineated by adopting a trained target area positioning model, training by adopting the following steps to obtain the target area positioning model:
acquiring a first sample set, and dividing the first sample set into a first training sample set and a first testing sample set; wherein each sample in the first set of samples comprises a first original medical image and a localization label of the first original medical image delineating a target region;
taking a pre-built neural network model as an initial target area positioning model;
training the initial target area positioning model based on the first training sample set until a first preset training end condition is met to obtain a candidate target area positioning model;
testing the candidate target area positioning model based on the first test sample set, and judging whether the test result meets a second preset training end condition: if so, taking the candidate target area positioning model as the trained target area positioning model;
if not, adjusting the model parameters of the candidate target area positioning model, and taking the candidate target area positioning model after parameter adjustment as the initial target area positioning model; and repeating the steps of training, testing and adjusting the model parameters of the candidate target area positioning model until the test result meets the second preset training end condition to obtain the trained target area positioning model.
Optionally, the localization label delineating the target region of each sample in the first set of samples is obtained by:
acquiring the first original medical image and a delineation target area of the first original medical image;
acquiring the position of the layer surface with the morphological difference exceeding a preset threshold value according to the delineation target area;
and dividing the first original medical image according to the position of the layer with the difference of the morphological information exceeding a preset threshold value to obtain the morphological distribution of the delineation target area, wherein the morphological distribution is used as a positioning label of the delineation target area of the first original medical image.
Optionally, the target segmentation model comprises a convolutional neural network model; before the trained target segmentation model is adopted to segment the medical image to be delineated, the method further comprises the following steps of training to obtain the target segmentation model:
acquiring a second sample set, and dividing the second sample set into a second training sample set and a second testing sample set; wherein each sample in the second set of samples comprises a second original medical image, a target label of the second original medical image, and multi-channel information corresponding to the target label;
taking a pre-built convolutional neural network model as an initial target area segmentation model;
training the initial target area segmentation model based on the second training sample set until a third preset training end condition is met to obtain a candidate target area segmentation model;
testing the candidate target segmentation model based on the second test sample set, and judging whether a test result meets a fourth preset training end condition, if so, taking the candidate target segmentation model as the trained target segmentation model;
if not, adjusting the model parameters of the candidate target area segmentation model, and taking the candidate target area segmentation model after the parameters are adjusted as the initial target area segmentation model; and repeating the steps of training, testing and adjusting the model parameters of the candidate target segmentation model until the test result meets the fourth preset training end condition to obtain the trained target segmentation model.
Optionally, before the pre-built convolutional neural network model is used as the initial target segmentation model, the method further includes:
calculating the number of morphological information classification categories of the target area delineated by any sample in the first sample set or calculating the number of morphological information classification categories of the multichannel information of the target area label of any sample in the second sample set;
taking the number of the classification categories as the number of channels of the convolutional neural network model;
and building and initializing the convolutional neural network model according to the number of channels of the convolutional neural network model.
Optionally, the method for automatically delineating the target region further includes performing a first pretreatment on each sample in the acquired first sample set, so as to obtain the target region localization model by using the first sample set after the first pretreatment;
and/or performing second preprocessing on each sample in the acquired second sample set to train the target region segmentation model by using the second preprocessed second sample set.
In order to achieve the above object, the present invention further provides an automatic target delineation apparatus for medical images, comprising:
the target area positioning module is configured to position a target area of a medical image to be sketched by adopting a trained target area positioning model so as to acquire multi-channel information of the target area, wherein the multi-channel information comprises morphological information and layer information corresponding to the morphological information;
and the target area segmentation module is configured to segment the medical image to be sketched by adopting a trained target area segmentation model according to the multi-channel information to obtain a target area sketching result of the medical image to be sketched.
In order to achieve the above object, the present invention also provides a medical image acquisition system comprising a medical image acquisition device, a processor and a memory, the memory having stored thereon a computer program;
the medical image acquisition device is configured to acquire a medical image to be sketched; when the computer program is executed by the processor, the target area of the medical image to be sketched is sketched by adopting the automatic target area sketching method of the medical image, and a target area sketching result of the medical image to be sketched is obtained.
Compared with the prior art, the method, the device and the system for automatically delineating the target area of the medical image have the following advantages:
according to the automatic target area delineation method for the medical image, provided by the invention, the target area of the medical image to be delineated is located by adopting the trained target area locating model, so that multi-channel information (including layer information and morphological information of the target area) of the target area is obtained. Therefore, the target area positioning model (such as a convolutional neural network model or a classification network model) with stronger edge form identification capability is adopted, so that the identification capability of target areas with different forms, such as the target areas with larger layer-to-layer deformation and complex boundary form structures, can be effectively improved, and good foundation can be laid for performing subsequent image segmentation quickly and accurately to obtain high-quality target area delineation results through the form information of the target areas. Furthermore, according to the multi-channel information, the medical image to be delineated is segmented by adopting a trained target segmentation model (such as a convolutional neural network model) to obtain a target delineation result of the medical image to be delineated, and because the target segmentation model can make up for the defect that the target segmentation model is not good at identifying fuzzy boundaries and images with larger morphological differences based on the constraint of morphological information (such as shape information in the morphological information, for example, one of a square and a rectangle) of the target, the target segmentation model can more quickly and accurately locate and segment the target (such as breast cancer, pelvic tumor and the like) with larger morphological differences. In summary, according to the automatic target delineation method for medical images provided by the present invention, the target positioning model is used to position the target and the target segmentation model is used to segment the target, so that advantages of the target positioning model and the target segmentation model are complemented and organically combined, the whole delineation process does not need manual participation, an end-to-end process is realized, and the delineation efficiency is high and the universality is strong; moreover, the problem of differentiation possibly caused by manual delineation can be reduced, and doctors can be better assisted to improve the quality and efficiency of radiotherapy plan design.
The device and the system for automatically delineating the target area of the medical image provided by the invention belong to the same inventive concept as the method for automatically delineating the target area of the medical image provided by the invention, so that the device and the system at least have all the advantages of the invention, and are not repeated.
Drawings
Fig. 1 is a schematic overall flowchart of a method for automatically delineating a target region of a medical image according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a training process for providing a target location model according to another embodiment of the present invention;
fig. 3 is a schematic diagram illustrating a training process of a target segmentation model according to a further embodiment of the present invention;
fig. 4 is a schematic structural diagram of an automatic target delineation apparatus for medical images according to a second embodiment of the present invention;
fig. 5 is a schematic structural diagram of a medical image acquisition system according to a third embodiment of the present invention;
wherein the reference numbers are as follows:
a target region positioning module-110, a target region segmentation module-120;
medical image acquisition equipment-210, imaging device-211, human-computer interaction device-212, examination table-213, processor-220, communication interface-230, memory-240, communication bus-250.
Detailed Description
The following describes the automatic target delineation method, device and system of medical images in detail with reference to the accompanying drawings. The advantages and features of the present invention will become more apparent from the following description. It should be noted that the drawings are in a very simplified form and are all drawn to a non-precise scale for the purpose of convenience and clarity only to aid in the description of the embodiments of the invention. To make the objects, features and advantages of the present invention comprehensible, reference is made to the accompanying drawings. It should be understood that the structures, proportions, sizes, and other elements shown in the drawings and described herein are illustrative only and are not intended to limit the scope of the invention, which is to be given the full breadth of the appended claims and any and all modifications, equivalents, and alternatives to those skilled in the art should be construed as falling within the spirit and scope of the invention. Specific design features of the invention disclosed herein, including, for example, specific dimensions, orientations, locations, and configurations, will be determined in part by the particular intended application and environment of use. In the embodiments described below, the same reference numerals are used in common for the same portions or portions having the same functions between different drawings, and the redundant description thereof may be omitted. In this specification, like reference numerals and letters are used to designate like items, and therefore, once an item is defined in one drawing, further discussion thereof is not required in subsequent drawings. Additionally, if the method described herein comprises a series of steps, and the order in which these steps are presented herein is not necessarily the only order in which these steps can be performed, some of the described steps may be omitted and/or some other steps not described herein may be added to the method.
It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrases "comprising a," "8230," "8230," or "comprising" does not exclude the presence of additional like elements in a process, method, article, or apparatus that comprises the element. The singular forms "a," "an," and "the" include plural referents, the term "or" is generally employed in a sense including "and/or," the terms "a number" and "an" are generally employed in a sense including "at least one," the terms "at least two" are generally employed in a sense including "two or more," and further, the terms "first," "second," and "third" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or to imply that the number of technical features indicated are all.
The core idea of the invention is that: aiming at the problems of low efficiency and poor consistency of the sketching results of the manual sketching of the target area by doctors in the prior art, the automatic sketching method, the automatic sketching device and the automatic sketching system for the target area of the medical image are provided, so that the whole sketching process does not need manual participation, an end-to-end process is realized, and the sketching efficiency is high and the universality is strong; meanwhile, the problem of differentiation possibly caused by human factors can be reduced, so that doctors can be better assisted to improve the quality and efficiency of radiotherapy plan design.
To achieve the above idea, the inventors of the present application have continuously found through intensive research and extensive practice that: based on the above research, the applicant of the present invention creatively proposes that the conventional image segmentation methods, such as edge detection, threshold segmentation or segmentation methods based on a spectral library, all require that the segmented object has significant boundary features, and the cancer target region is a region where the cancer focus and the surrounding normal tissue are infiltrated and invaded by each other, and often has no definite boundary information, which results in the defect of poor segmentation accuracy of the conventional method: the method comprises the steps of firstly acquiring multi-channel information of a target area of a medical image to be sketched by adopting a target area positioning model, and then segmenting the medical image to be sketched by adopting a target area segmentation model according to the acquired multi-channel information of the target area, so that segmentation efficiency and segmentation precision of the target area segmentation model can be remarkably improved by organically combining the target area positioning model and the target area segmentation model.
The multi-channel information may include morphological information of the target, for example, the target is divided into at least two types according to morphological distribution of the target, and then the two types of morphology are processed by using a target segmentation model to obtain a target delineation result. It should be noted that the method for automatically delineating the target area of the medical image provided by the present invention is particularly suitable for delineating the target area with large layer-to-layer deformation and complex boundary morphological structure, including but not limited to CTV after radical treatment of breast cancer, pelvic tumor CTV, etc. Further, the method may be performed by a target region automatic delineation apparatus for medical images or a medical image acquisition system provided in the present invention, where the target region automatic delineation apparatus may be implemented by software and/or hardware, and the target region automatic delineation apparatus may be deployed on the medical image acquisition system. The medical image acquisition system can be a personal computer, a mobile terminal and the like, and the mobile terminal can be a mobile phone, a tablet personal computer and other hardware equipment with various operating systems.
The following describes the method, device and system for automatically delineating the target region of the medical image provided by the invention in detail respectively.
Example one
The embodiment provides a target area automatic delineation method of a medical image. Specifically, please refer to fig. 1, which schematically shows a flowchart of the method for automatically delineating the target region of the medical image provided by this embodiment, and as can be seen from fig. 1, the method for automatically delineating the target region of the medical image provided by this embodiment includes the following steps:
s100: positioning a target area of a medical image to be delineated by adopting a trained target area positioning model to acquire multi-channel information of the target area, wherein the multi-channel information comprises morphological information and layer information corresponding to the morphological information;
s200: and according to the multi-channel information, segmenting the medical image to be sketched by adopting a trained target region segmentation model to obtain a target region sketching result of the medical image to be sketched.
With such configuration, the target region of the medical image to be delineated is positioned by adopting the trained target region positioning model, so as to obtain multi-channel information (including layer information and morphological information of the target region) of the target region. Therefore, the target area positioning model (such as a convolutional neural network model or a classification network model) with stronger edge form identification capability is adopted, so that the identification capability of target areas with different forms, such as the target areas with larger layer-to-layer deformation and complex boundary form structures, can be effectively improved, and good foundation can be laid for performing subsequent image segmentation quickly and accurately to obtain high-quality target area delineation results through the form information of the target areas. Furthermore, according to the multi-channel information, the medical image to be delineated is segmented by adopting a trained target segmentation model (such as a convolutional neural network model) to obtain a target delineation result of the medical image to be delineated, and because the target segmentation model can make up for the defect that the target segmentation model is not good at identifying fuzzy boundaries and images with larger morphological differences based on the constraint of morphological information (such as shape information in the morphological information, for example, one of a square and a rectangle) of the target, the target segmentation model can more quickly and accurately locate and segment the target (such as breast cancer, pelvic tumor and the like) with larger morphological differences. In summary, according to the automatic target delineation method for medical images provided by the present invention, the target positioning model is used to position the target and the target segmentation model is used to segment the target, so that advantages of the target positioning model and the target segmentation model are complemented and organically combined, the whole delineation process does not need manual participation, an end-to-end process is realized, and the delineation efficiency is high and the universality is strong; moreover, the problem of differentiation possibly caused by manual delineation can be reduced, and doctors can be better assisted to improve the quality and efficiency of radiotherapy plan design.
It should be noted that, those skilled in the art should understand that the present invention does not limit the medical image to be sketched. Firstly, the obtaining mode of the medical image to be outlined is not limited, and the medical image to be outlined may be a medical image obtained by real-time scanning of a medical imaging system, or a medical image obtained by reading from a cloud, a server, a storage medium, and the like through a network access, a data reading, and the like. Secondly, the Imaging mode/format of the medical image to be outlined is not limited, for example, the medical image to be outlined may be a PET (Positron Emission Tomography) image, a CT (Computed Tomography) image, an MRI (Magnetic Resonance Imaging) image, a CBCT (cone beam Computed Tomography) image, etc., and in other embodiments, the medical image to be outlined may be a DICOM (Digital Imaging and Communications in Medicine) image.
It should be noted that the present invention does not limit the type of the target region and the organ tissues where the target region is located. The target volume may include various types of tumor radiotherapy target volumes, including, for example, but not limited to, a breast tumor target volume, a rectal tumor target volume, a cervical tumor target volume, a thyroid tumor target volume, a pancreatic tumor target volume, and the like. It will be understood by those skilled in the art that the target area is exemplified herein, and not limited to the scope of the target area, but is merely illustrative. More specifically, the target region of the medical image to be delineated is a part where a lesion (such as a tumor) is located, and the target region delineating result is that the target region is based on a segmented image of the medical image to be delineated.
Further, in a preferred embodiment, before the target region positioning model is trained to position the medical image to be delineated, positioning preprocessing is further performed on the medical image to be delineated. Specifically, in some of these embodiments, the positioning pre-processing comprises: the medical image to be delineated (for example, an original medical image including the rectum of the patient obtained by scanning the abdominal region of the patient with a medical imaging system) is subjected to image segmentation by using an image segmentation method known to those skilled in the art to obtain an image of a region of interest (for example, a rectal region where a target region of a rectal tumor is located), and the segmented region of interest image is used as the medical image to be delineated. Further, in some other embodiments, the positioning preprocessing further includes, but is not limited to, resampling and/or denoising the medical image to be delineated, and the like. In some embodiments, the medical image to be delineated may be sampled according to a preset resolution, so as to obtain a resampled medical image to be delineated, which satisfies the input resolution condition of the target area positioning model. Further, a gaussian filter may be further used to filter noise information in the medical image to be outlined, which obviously is not limited in this respect, and in some other embodiments, other common filters may also be used to perform denoising processing on the medical image to be outlined. Therefore, the noise information in the medical image to be segmented is filtered, so that the image quality of the medical image to be sketched can be effectively improved, and a good foundation is laid for obtaining a high-quality target region sketching result subsequently.
Correspondingly, adopt the target area location model that trains well to treat the target area of delineating medical image and fix a position to acquire the multichannel information of target area, include: and positioning the target area of the medical image to be delineated after positioning pretreatment by adopting the trained target area positioning model so as to obtain multi-channel information of the target area.
In one exemplary embodiment, the positioning a target region of a medical image to be delineated by using a trained target region positioning model to obtain multi-channel information of the target region includes:
acquiring the morphological distribution of the target area by adopting the target area positioning model according to the slice image of the medical image to be sketched;
obtaining multi-channel information of the target area according to the morphological distribution of the target area; the multichannel information comprises a plurality of pieces of channel information, and each piece of channel information comprises morphological information and layer information corresponding to the morphological information.
The invention provides a target area automatic delineation method of a medical image, which comprises the steps of firstly adopting a target area positioning model to obtain the morphological distribution of a target area, and then obtaining multi-channel information of the target area according to the morphological distribution of the target area. This can improve the positioning accuracy of the target region.
Specifically, in some embodiments, the morphological distribution is morphological information divided by the target area localization model and slice information corresponding to the morphological information, and then the morphological information is classified, and the multi-channel information is obtained according to a classification result. In other embodiments, the target region localization model may directly obtain multi-channel information, that is, the morphological distribution in this case is multi-channel information obtained by classifying the morphological information.
It should be noted that the medical image to be outlined is a three-dimensional medical image. From the above disclosure, those skilled in the art will understand that the target region localization model is inputted into the two-dimensional tomographic image of the medical image to be delineated. More specifically, the medical images to be delineated are preferably MR images or CT images, and each medical image to be delineated (such as a CT image) may include a plurality of consecutive slice images (sequence images). The slice images are medical images of different sections of organ tissues of a patient, for example, the slice images are two-dimensional medical images of a breast at different axial positions along the head-foot axial direction.
Further, the slice information includes a slice position of a slice image containing target region information corresponding to morphological information when the slice has the target region information; the morphological information includes, but is not limited to, a geometric shape, including, but not limited to, one of a rectangle, a circle, and/or a square, a location, and/or a pixel value, etc., of the target region in image coordinates of each slice image. Preferably, the morphological information comprises a geometry of the target volume.
In one exemplary embodiment, the method for automatically delineating the target region of the medical image further includes: and performing post-processing on the initial target area image obtained by segmenting the medical image to be sketched, and taking the post-processed initial target area image as the target area sketching result.
Specifically, the post-processing includes, but is not limited to: and carrying out operations such as hole filling on the initial target area image, acquiring and removing noise points from the maximum communication region of the initial target area image after the acquired holes are filled, and taking the maximum communication region of the initial target area image after the holes are filled and the noise points are removed as the target area delineation result.
It should be noted that, under the influence of the segmentation precision of the target region segmentation model, an interference region may exist on the initial target region image, and the area of the interference region should be smaller than the area of the target region delineation result theoretically, so that the maximum connected region is obtained by analyzing the connected region of the initial target region image, and the influence of the interference region can be effectively avoided. Therefore, according to the automatic target area delineation method for the medical image, provided by the invention, the maximum connected domain of the initial target area image after the holes are filled and the noise points are removed is used as the delineation result of the target area, so that the delineation quality of the target area image can be further improved, and a doctor is further assisted to improve the quality and efficiency of radiotherapy plan design. It is noted that, as will be appreciated by those skilled in the art, the initial target region image may be hole-filled using a method including a hole-filling method known to those skilled in the art, such as a morphological close-up operation of dilation-followed-by-erosion, and will not be further described herein.
Preferably, in one exemplary embodiment, the target region localization model is a convolutional neural network model. It should be noted that, besides the convolutional neural network, the target area positioning model may also adopt a classification network model or use a machine learning method such as random forest, XGoost, adaboost, and the like. The invention is not limited in this regard. Preferably, the target area positioning model may adopt an encoder of a unet network and a classification network constructed by adding a full connection layer after the encoder.
Preferably, in one exemplary embodiment, before the target region to be delineated on the medical image is located by using the trained target region location model, the method further includes training the target region location model by using the following steps. Specifically, referring to fig. 2, as can be seen from fig. 2, the training step of the target area localization model includes:
a1: acquiring a first sample set, and dividing the first sample set into a first training sample set and a first testing sample set; wherein each sample in the first set of samples comprises a first original medical image and a localization label of the first original medical image delineating a target region;
a2: taking a pre-built convolutional neural network model as an initial target area positioning model;
a3: training the initial target area positioning model based on the first training sample set until a first preset training end condition is met to obtain a candidate target area positioning model;
a4: testing the candidate target area positioning model based on the first test sample set, and judging whether the test result meets a second preset training end condition: if yes, executing the step A5; if not, adjusting model parameters of the candidate target area positioning model, taking the candidate target area positioning model after parameter adjustment as the initial target area positioning model, and executing the step A3;
a5: and using the candidate target area positioning model as the trained target area positioning model.
Therefore, according to the automatic target area delineation method for the medical image, provided by the invention, the target area positioning model is obtained by training a pre-constructed convolutional neural network model in a manner of combining training and testing, so that the accuracy of the obtained multi-channel information of the target area can be improved, and a foundation is laid for more efficiently and accurately segmenting the target area segmentation model to obtain the target area delineation result.
Specifically, the first original medical image of each sample in the first sample set should be a medical image with a completed target region of a different medical image of the same tissue organ, that is, a three-dimensional medical image (such as a CT image, an MR image, etc.) that a doctor or an expert has completed target region delineation on the same diseased region of a different patient or the same patient at different acquisition times, for example, a CT image with a completed target region delineation on 100 patients with breast diseases, a CT image with a completed target region delineation 100 patients with rectum diseases, or a CT image with a completed target region delineation on 100 patients with cervix diseases.
Further, in some embodiments, the positioning tag corresponding to the delineation target area of the first original medical image is bedding information and morphological information where the delineation target area of the diseased part is located. More specifically, since the two-dimensional tomographic image (i.e., slice image) of the first original medical image is input when the target region localization model is trained, the localization label includes a slice image layer position of the two-dimensional tomographic image of the first original medical image that delineates the target region, and morphological information (including, but not limited to, a shape, an area, a contour, a position, etc. of the target region on the slice image layer) of the target region on the slice image layer. In some other embodiments, the positioning tag may be multi-channel information classified according to the morphological information. The first original medical image is a medical image which is already delineated by the target area, so that the positioning label corresponding to the delineated target area of the first original medical image can be obtained more conveniently. In addition, the method for acquiring the positioning tag of the target region of the first original medical image will be described in detail below, and will not be described herein.
Specifically, the candidate target localization model is tested by a first test sample set to evaluate the algorithm accuracy of the candidate target localization model. Thereby, the accuracy of the target localization model is further improved. In particular, in some of these embodiments, the first preset training end condition may be that the training of the candidate target localization model is stopped when the loss function values on the first training sample set do not decrease for a plurality of rounds. Further, in some other embodiments, the second preset training end condition includes that the output accuracy is less than the preset accuracy. For example, if the output accuracy is smaller than the preset accuracy, the candidate target area positioning model is continuously trained. It should be noted that, the preset accuracy is not limited in the present invention, and the preset accuracy is set according to specific situations, for example, the preset accuracy is set to 95%.
Further, the first sample set may be divided into a first training sample set and a first testing sample set according to a certain proportion. It should be noted that, the present invention is not limited to the ratio of the first training sample set and the first testing sample set, and the ratio of the first training sample set and the first testing sample set may be set according to specific situations, for example, 75% of samples in the first training sample set may be selected as the first training sample set, and the remaining 25% of samples may be selected as the first testing sample set, for example, 150 samples out of 200 samples may be used for training a model, and the other 50 samples may be used for testing the model. Further, in some embodiments, when the output accuracy is less than the preset accuracy, the initial target region localization model may be retrained by increasing the number of samples in the first sample set or readjusting the ratio of the first training sample set and the first testing sample set until the output accuracy is greater than or equal to the preset accuracy. Still further, those skilled in the art will appreciate that in other embodiments, the first sample set may be divided into the first training sample set, the first testing sample set, and the first verification sample set according to a certain proportion. The first training sample set is used for training, the first testing sample set is used for testing, and the first verifying sample set is used for verifying, so that the algorithm precision of the target area positioning model is further improved. Similarly, the proportions of the first training sample set, the first testing sample set and the first verification sample set may be set reasonably according to specific situations, which is not limited in the present invention.
In addition, as will be understood by those skilled in the art, the parameters of the neural network model include two categories: characteristic parameters and hyper-parameters. The characteristic parameters are continuously and iteratively learned by a neural network model and are used for learning image characteristics, such as the geometric shape, texture, position and other characteristics of the target area in the first original medical image. The characteristic parameters include a weight parameter and a bias parameter. The hyper-parameters are parameters manually set during training, and the characteristic parameters can be learned from the sample only by setting the proper hyper-parameters. The hyper-parameters may include a learning rate (which may be considered as a step size), the number of hidden layers, the size of the convolution kernel, the number of training iterations, and the batch size per iteration. Because the training process of the model is actually the process of minimizing the loss function, and the derivation can quickly and simply achieve the goal, the derivation method is a gradient descent method, so the invention can preferentially adopt the gradient descent method to update the parameters of the deep neural network model until the error requirement (the loss function value is not changed) is converged or the set end (such as a certain number of times) condition is reached, and the network parameters at the moment are stored. That is, the first preset training end condition may be that the loss function value converges to the error requirement or that a preset end condition is reached. It should be noted that, the present invention does not limit the specific values of the hyper-parameters, and the values should be set reasonably according to actual needs. In addition, the loss function for training the target region localization model is not particularly limited, for example, the loss function may be a cross entropy function.
Preferably, in one exemplary embodiment, the target region automatic delineation method further includes: performing a first pre-treatment on each sample in the acquired first sample set to obtain the target area localization model by using the first pre-treated sample set.
Since the data of the first original training sample is limited, and deep learning needs to be performed on certain data to have certain robustness, in order to increase the robustness, the first preprocessing includes a data amplification operation. And the data amplification is used for improving the data capacity so as to increase the generalization capability of the deep neural network model and further improve the model positioning effect. Further, the first preprocessing may further include performing denoising processing on each sample in the first sample set after the data amplification, so as to remove noise in the image and improve the image quality of the first sample set.
In particular, in some of these embodiments, the data augmentation operation may be performed by randomly rigidly transforming the first raw medical image, including in particular: rotation, scaling, translation, flipping, and grayscale transformation. More specifically, the first original medical image may be translated by-10 to 10 pixels, rotated by-30 ° to 30 °, horizontally flipped, vertically flipped, scaled by 0.8 to 1.2 times, grayscale transformed, etc. to accomplish data augmentation of the medical image. By the transformation, the original 20 images can be expanded to 2000 images, 1600 images can be used for model training, and the rest 400 images can be used for model testing. In other embodiments, the denoising process includes at least one of region-of-interest cropping, organ tissue segmentation, resampling, etc., of the first raw medical image. For example, the first original medical image may include more background areas, which results in more resources being wasted in subsequent network training, and thus image cropping is required. For example, the first original medical image with the size of 512 × 512 is reduced to 256 × 256 to remove the background region and only the region where the organ tissue is located is reserved, so that the requirement of the network model on the image size can be used, and the training efficiency of the model can be improved. The segmentation and resampling of the organ tissue may be performed by techniques known to those skilled in the art, and will not be described herein.
Preferably, in one of the exemplary embodiments, the localization label delineating the target region of each sample in the first sample set is obtained by:
acquiring the first original medical image and a delineation target area of the first original medical image;
acquiring the position of the layer surface with the morphological difference exceeding a preset threshold value according to the delineation target area;
and dividing the first original medical image according to the layer position of which the morphological difference exceeds a preset threshold value to obtain the morphological distribution of the delineation target area so as to serve as a positioning label of the delineation target area of the first original medical image.
The invention provides an automatic target area delineating method of a medical image, which comprises the steps of firstly obtaining a delineation target area of a first original medical image, and then obtaining a layer position with a morphological difference exceeding a preset threshold value according to the delineation target area; and then, dividing the first original medical image according to the position of the layer where the morphological difference exceeds a preset threshold value to obtain the morphological distribution of the delineation target area, wherein the morphological distribution is used as a positioning label of the delineation target area of the first original medical image. Namely, the positioning label of the first original medical image which delineates the target area comprises morphological information which delineates the target area and the layer position corresponding to each piece of morphological information. Therefore, the identification capability of the target area with large layer-to-layer deformation and complex boundary morphological structure can be further effectively improved, and a good foundation is laid for obtaining a high-quality target area delineation result by quickly and accurately performing image segmentation subsequently through morphological information (such as shape information) delineating the target area. Preferably, the target region of the first original medical image is a target region delineation image of the diseased part of a doctor or an expert.
Specifically, in some embodiments, after the first original medical image and the delineation target area of the first original medical image are acquired, the position of a slice whose morphological difference exceeds a preset threshold and the morphological information of the delineation target area at the slice can be obtained by positioning the delineation target area; then, according to the position of the layer where the morphological difference exceeds a preset threshold and the morphological information of the delineation target area at the layer, the first original medical image is divided according to the position of the layer where the morphological difference exceeds the preset threshold, and the morphological distribution of the delineation target area is obtained and used as a positioning label of the delineation target area of the first original medical image.
It should be noted that, as can be understood by those skilled in the art, when the delineation target region is located, the present invention does not limit the specific method for acquiring the slice positions with morphological differences exceeding a preset threshold and the morphological information corresponding to the delineation target region. Specifically, in some embodiments, an algorithm may be used to locate the slice position where the delineation target region exceeds a preset threshold (with a large geometric deformation), so as to obtain the slice position of the multi-channel information; in other embodiments, the level position may be derived from bony landmarks. And then according to the bedding information, obtaining the morphological information of the delineation target area by adopting common Pruker superposition, fourier transformation, characteristic shape analysis, enhanced characteristic shape analysis and the like in morphology measurement. Specifically, taking the breast cancer target area as an example: the breast cancer target area is generally greatly deformed in the head-foot axis direction, and in some embodiments, an algorithm may be used to locate the level position where the breast cancer target area exceeds a preset threshold. For example, if the area difference between two adjacent slice images exceeds 30%, the morphological difference between the two slice images is considered to be large; or the morphology difference is large when the coincidence area of two adjacent slice images is less than 60%. In other embodiments, the level information of the breast cancer target region can also be obtained according to the position relationship between the breast cancer target region and the breast angle bone. Further, those skilled in the art should understand that the above is only an exemplary illustration of the preferred embodiment, and in practical applications, not only the geometric shape, but also other features with significant differences can be used to obtain the slice information and the morphological information, such as texture features in the slice mask region (distribution of CT values of images in the region, gray level co-occurrence matrix, etc.), or medical prior information, including but not limited to the position information of bony landmarks, etc.
Specifically, for example, whether the geometric shape of the morphological information that delineates the target region exceeds a preset threshold value or not is used to position the position of the slice, the geometric shape of the morphological information is first obtained, for example, 100 slices exist in a certain first original medical image, and the target region is not delineated in the 1 st layer to 20 th layer; the geometric shapes of the morphological information of the delineation target areas at the 21 st layer to the 30 th layer are squares, the geometric shapes of the morphological information of the delineation target areas at the 31 st layer to the 70 th layer are circles, and the geometric shapes of the morphological information of the delineation target areas at the 71 st layer to the 100 th layer are triangles, so that the positions of the layers with larger morphological differences are respectively as follows: layer 20, layer 30 and layer 70, the localization label of the first original medical image is: the bedding surface position is 21 st to 30 th layers, and the morphological information is a square; the layer surface positions are 31 th to 70 th layers, and the morphological information is circular; the bedding positions are layers 71-100, and the morphological information is a triangle. Correspondingly, the morphological information is classified, and according to the classification result, a plurality of pieces of channel information in the multi-channel information of the target area are obtained, where the number of the pieces of channel information is 3 (corresponding to a square, a circle, and a triangle, respectively).
It should be noted that, similar to the medical image to be sketched, the first original medical image is also not limited in the present invention, and for more detailed contents, reference is made to the related description of the medical image to be sketched, which is not repeated herein. Further, the invention also does not limit the manner of obtaining the target region of the first original medical image, for example, in some embodiments, the target region of the first original medical image may be a gold standard delineation image of the target region of the diseased region; in other embodiments, the target area of the diseased part can be manually delineated by a doctor; in some embodiments, the first original medical image may be obtained by inputting the first original medical image into the target delineation model and evaluating the target delineation model by the user, and the like, which are not specifically limited herein.
In addition, it should be understood by those skilled in the art that, although the description above has been given by taking the example of positioning the delineation target area, obtaining the slice information whose morphological difference exceeds the preset threshold and the morphological information corresponding to the delineation target area, and taking the slice information and the morphological information as the positioning label corresponding to the delineation target area of the first original medical image. However, the present invention does not limit the specific acquisition manner of the positioning tag corresponding to the delineation target region of the first original medical image. For example, in some embodiments, the first original medical image may be based on a manual delineation of a target region of the first original medical image; in other embodiments, the first original medical image may be obtained by inputting the delineation target area of the first original medical image into the shape recognition model and evaluating the shape recognition model by the user; in other embodiments, the positioning label corresponding to the delineation target area may also be obtained through a mathematical algorithm, and the like, which is not specifically limited herein.
In one exemplary embodiment, the target volume segmentation model includes a convolutional neural network model. Specifically, a ResNet50 neural network model in the prior art may be adopted as the target region segmentation model. Because the ResNet50 uses jump connection (or called shortcut), it directly transmits the network layer activation value of a certain layer to the deeper layer of the network, in addition, the jump connection only transmits data, through the jump connection, the signal can be transmitted without attenuation when reversely transmitting, and it is not necessary to worry about the change of the gradient, and it can transmit effective gradient to the upper layer, therefore, the problem of gradient disappearance caused by deepening the network layer can be effectively relieved through the jump connection, through the stacking of Residual block, it can construct very deep network model, and make the deep network layer also carry on effective training. For the specific structure of the ResNet50 neural network model, reference may be made to the prior art, and details thereof are not repeated here. Of course, as will be understood by those skilled in the art, the target segmentation model may also adopt other convolutional neural network models in the prior art, such as one of the Unet network model, the transform model and the GAN model or other convolutional neural network models in addition thereto, which is not limited by the present invention.
Preferably, in one exemplary embodiment, before segmenting the medical image to be delineated by using the trained target segmentation model, the method further includes training the target segmentation model by using the following steps. In particular, please refer to fig. 3, which schematically shows the training step of the target segmentation model. As can be seen from fig. 3, the training step of the target segmentation model includes:
b1: acquiring a second sample set, and dividing the second sample set into a second training sample set and a second testing sample set; wherein each sample in the second sample set comprises a second original medical image, a target label of the second original medical image, and multi-channel information corresponding to the target label;
b2: taking a pre-built convolutional neural network model as an initial target area segmentation model;
b3: training the initial target segmentation model based on the second training sample set until a third preset training end condition is met to obtain a candidate target segmentation model;
b4: testing the candidate target area segmentation model based on the second test sample set, and judging whether the test result meets a fourth preset training end condition, if so, executing the step B5; if not, adjusting model parameters of the candidate target area segmentation model, taking the candidate target area segmentation model after parameter adjustment as the initial target area segmentation model, and executing the step B3;
b5: and taking the candidate target segmentation model as the trained target segmentation model.
With the configuration, according to the automatic target delineation method for the medical image, provided by the invention, the target segmentation model is obtained by training a pre-constructed convolutional neural network model in a manner of combining training and testing, and the multi-channel information corresponding to the target label is obtained based on the trained target positioning model, so that the accuracy and the delineation efficiency of the target delineation result can be further improved.
It is specifically noted that the present invention is not limited to a particular source of the second training sample, as will be appreciated by those skilled in the art. However, the second original medical image of each sample in the second sample set should not only be a medical image delineating a completed target region of a different medical image located in the same tissue organ, but also correspond to the same tissue organ as the first original medical image of each sample in the first sample set. I.e. three-dimensional medical images (such as CT images, MR images, etc.) that a doctor or a specialist has already completed target delineation for different patients or the same patient at different acquisition times for the same affected part, for example, if the first sample set is a CT image completed target delineation for 100 breast affected patients, then the second sample set should also be a CT image completed target delineation for several breast affected patients; if the first sample set is a CT image of a complete target delineation of 100 patients with rectal disease, then the second sample set should also be a CT image of a complete target delineation of several or patients with rectal disease. By analogy, the above description is not repeated. Optimally, the second original medical image is a three-dimensional medical image obtained after the slice image of each sample in the first sample set is subjected to three-dimensional reconstruction, and the three-dimensional medical image subjected to three-dimensional reconstruction is used for target region segmentation, so that the continuity between layers in the segmented target region can be ensured.
Further, in some embodiments, the target region label of the second original medical image is preferably a target region delineation image of the diseased region by a physician, and the multi-channel information corresponding to the target region label of the second original medical image may be obtained by inputting the second training sample into the target region localization model for recognition, or may be obtained by other algorithms or manual classification by the physician. Obviously, the present invention does not limit the manner of obtaining the multi-channel information corresponding to the target label of the second original medical image, and in some other embodiments, the multi-channel information corresponding to the target label of the second original medical image may also be obtained by other recognition algorithms or even manually sketching. Further, the present invention does not limit the first sample set and the second sample set, and a first original medical image in the first sample set may be the same as or different from a second original medical image in the second sample set, but the first original medical image and the second original medical image should be images of the same diseased part. Further, similar to the training of the target area localization model, the present invention also does not limit the allocation ratio of the second training sample set and the second testing sample set in the second sample set, for more detailed description, please refer to the adaptive understanding of the distribution ratio of the first training sample set and the first testing sample set in the first sample set in the target area localization model, and for avoiding redundant description, the description is not repeated here.
In addition, similar to the training of the target region localization model, the candidate target region localization model is tested by the second test sample set to evaluate the algorithm accuracy of the candidate target region segmentation model. This further improves the accuracy of the target segmentation model. Specifically, in some of these embodiments, the fourth preset training end condition includes that the output accuracy is less than the preset accuracy. For example, if the output accuracy is smaller than the preset accuracy, the candidate target segmentation model is continuously trained. It should be noted that, the preset accuracy is not limited, and is set according to specific situations, for example, the preset accuracy is set to 95%. In some other embodiments, the third preset training end condition may also be that the training of the candidate target localization model is stopped when the loss function value on the second training sample set does not decrease for a plurality of rounds. Similarly, the loss function for training the target segmentation model is also not limited in the present invention, for example, a cross entropy function and a DICE function may each take 50% weight as the loss function of the target segmentation model. For further details, please refer to the related description in the above-mentioned training method of the target area positioning model, and the description is not repeated here.
Preferably, in one preferred embodiment, before taking the pre-built convolutional neural network model as the initial target segmentation model, the method further includes:
calculating the number of classification categories of morphological information of the target area delineated by any sample in the first sample set or calculating the number of classification categories of morphological information of the multi-channel information of the target area label of any sample in the second sample set;
taking the number of the classification categories as the number of channels of the convolutional neural network model;
and building and initializing the convolutional neural network model according to the number of channels of the convolutional neural network model.
With such configuration, the classification number is used as the number of channels of the convolutional neural network model (i.e., the initial target segmentation model), so that the trained target segmentation model can quickly and accurately identify the target of the medical image to be delineated according to the specific shape (geometric model) of the positioning label, thereby better segmenting the target.
It should be noted that the classification category of the geometric shape may be set as needed. For example: in some embodiments, the geometric shape of the morphological information includes four types, i.e., a circle, a square, a rectangle, and an ellipse, and the classification type of the circle is defined as a first type, the classification type of the square is defined as a second type, the classification type of the rectangle is defined as a third type, and the classification type of the ellipse is defined as a fourth type. There are 4 classification categories, and accordingly, the number of channels of the target segmentation model can be set to 4. In some other embodiments, circular and elliptical classification categories may be used as the first category, square classification categories as the second category, and rectangular classification categories as the third category. There are 3 classification categories, and accordingly, the number of channels of the target segmentation model can be set to 3. More specifically, taking the classification categories of the geometric shape of the morphological information of the target area label of the multi-channel information of all the samples in the second sample set, the number of the classification categories is calculated as follows: in one embodiment, if the second training sample set has 1000 samples, 1 sample is selected from the 1000 samples (since each sample in the second sample set is for a medical image of the same tissue and organ, morphological information of a target region of the same diseased type based on the same organ and tissue is substantially similar, and thus even if one sample is selected from the sample set, the sample is reliable), if the sample has 100 (obviously, this is only an idealized exemplary description and does not represent an actual application scenario) slice images, wherein the geometric shape of the morphological information of 60 slice images in total is circular in the 100 slice images, the geometric shape of the morphological information of 20 slice images is rectangular, the geometric shape of the morphological information of 20 slice images is square, and if the classification category of the circular is used as the first category, the classification category of the rectangular is used as the second category, the classification category of the square is used as the third category, and the number of the classification categories is 3.
Further, preferably, before classifying the information in the positioning tag of each sample in the second training sample set according to the geometric shape of the positioning tag and acquiring the classification number, the shape of the positioning tag may be further preprocessed. For example, irregular geometric shapes between squares and circles may be classified as squares or circles according to a predetermined error criterion.
Preferably, in one exemplary embodiment, each sample in the acquired second sample set is subjected to a second preprocessing, so as to train the target segmentation model by using the second sample set after the second preprocessing.
In particular, similar to the first pre-processing, the second pre-processing may also include data amplification and de-noising processing to increase the robustness of the target segmentation model and the image quality of the second sample set. Further, the second preprocessing may further include a normalization processing. Preferably, the first preprocessed first sample set may be used as the second sample set. In other words, the normalization process may be performed on the basis of the first pre-process, or may be performed without the first pre-process. It should be understood by those skilled in the art that the present invention does not limit the specific steps of the second preprocessing, i.e. the sequence of the data amplification, denoising and normalization processing is not limited.
Therefore, according to the automatic target area delineation method for the medical image, provided by the invention, each sample in the second sample set is subjected to second pretreatment, and the second sample set subjected to the second pretreatment is used for training to obtain the target area segmentation model, so that the gradient can be effectively prevented from generating large change in the training process of the network model, and a neural network has a better iterative convergence effect.
More specifically, the normalization may further include adjustment of a target size of the sample image and truncation processing of the pixel value. First, due to differences between different medical imaging devices (CT scanning devices, such as MRI scanning devices), the scanning parameters of the second original medical image in the second sample set acquired by different medical imaging devices may be different, resulting in different original image sizes of the samples, and thus, the second original medical image in each sample and the target area label image of the second original medical image need to be scaled to adjust to the target size required by the initial target area segmentation model. Specifically, the second original medical image may be adjusted to a target size by spline interpolation (trilinear interpolation), and the target area label image of the second original medical image may be adjusted to a target size by nearest neighbor interpolation. Of course, other interpolation methods may also be employed to adjust the second original medical image and the target area label image of the second original medical image to a target size, as will be appreciated by those skilled in the art. Here, there is no one-to-one listing. Further, the truncation processing of the pixel values includes defining the pixel values in the second original medical image scaled to the target size and the target area label image of the second original medical image to be within a certain pixel value range. For example, in one exemplary embodiment, the pixel value of the pixel point with the pixel value less than 100 in the second original medical image may be set to 100, the pixel value of the pixel point with the pixel value greater than 800 may be set to 800, and the pixel value of the pixel point with the pixel value in the range of [100,800] may be kept unchanged.
Example two
The present embodiment provides an automatic target area delineating device for a medical image, and specifically, please refer to fig. 4, which schematically shows a structural schematic diagram of the automatic target area delineating device for a medical image provided by the present embodiment. As can be seen from fig. 4, the present embodiment provides an apparatus for automatically delineating a target region of a medical image, which includes a target region positioning module 110 and a target region segmentation module 120.
Specifically, the target positioning module 110 is configured to position a target of the medical image to be delineated by using the trained target positioning model to obtain multi-channel information of the target, where the multi-channel information includes morphological information and layer information corresponding to the morphological information. The target area segmentation module 120 is configured to segment the medical image to be delineated by using a trained target area segmentation model according to the multi-channel information, so as to obtain a target area delineation result of the medical image to be delineated.
Since the apparatus for automatically delineating a target region of a medical image provided in this embodiment and the method for automatically delineating a target region of a medical image provided in the first embodiment belong to the same inventive concept, at least all advantages of the method for automatically delineating a target region of a medical image are provided.
EXAMPLE III
Fig. 5 schematically shows a structural schematic diagram of the medical image acquisition system provided in this embodiment. As shown in fig. 5, the present embodiment provides a medical image acquisition system including a medical image acquisition device 210, a processor 220, and a memory 240. The memory 240 has stored thereon a computer program, wherein the medical image acquisition device 210 is configured to acquire a medical image to be delineated; when the computer program is executed by the processor 220, the target area of the medical image to be sketched is sketched by using the target area automatic sketching method of the medical image according to any implementation manner in the first embodiment, so as to obtain a target area sketching result of the medical image to be sketched.
Since the medical image acquisition system provided in this embodiment and the method for automatically delineating the target area of the medical image provided in the first embodiment belong to the same inventive concept, the medical image acquisition system provided in this embodiment at least has all the advantages of the method for automatically delineating the target area of the medical image, and for avoiding redundancy, no description is provided here, and for more detailed contents, reference is made to the related description of the method for automatically delineating the target area of the medical image.
It is specifically noted that, as will be appreciated by those skilled in the art, the medical image acquisition apparatus 210 is a medical image acquisition system for acquiring medical images, and in some embodiments, the medical image acquisition apparatus 210 includes, but is not limited to, at least one imaging device 211. For example, the imaging device 211 may be one or more of a Computed Radiography (CR), digital Radiography (DR), computed Tomography (CT), magnetic Resonance Imaging (MRI), digital Subtraction Angiography (DSA), emission Computed Tomography (ECT), positron Emission Tomography (PET), or other similar imaging apparatus. Further, the medical image acquisition apparatus 210 may further include a human-computer interaction device 212 and an examination table 213, and the human-computer interaction device 212 may implement interaction and information exchange between the medical image acquisition system and the user. For example, human interaction device 212 includes one or more input/output devices. The input/output device may include a keyboard, a mouse, a display device, or a touch screen, etc. For example, a doctor or a patient may enter information relating to the patient in an input device, or information relating to a subsequent scan. For another example, the display device may be configured to provide one or more user interfaces to a user, which may include a plurality of interface elements with which the user may interact, etc.; the medical image to be delineated and the target volume delineation result (target volume segmentation image) can be imaged to the user through the interface element. For example, the examination couch 213 is used for carrying a patient to more conveniently acquire medical images of a portion of the patient to be scanned, such as the chest, abdomen, upper limbs, lower limbs, etc. of the patient. According to the user instruction received by the human-computer interaction device 212, the main operation program of the medical image acquisition system controls the imaging device 211 and the examining table 213 to acquire a medical image, and controls the processor 220 to execute the computer program, and the target region of the medical image to be sketched is sketched by using the target region automatic sketching method of the medical image according to any one of the embodiments, so as to obtain a target region sketching result of the medical image to be sketched. In some other embodiments, the medical image capturing device 210 may also capture medical images from a PACS system or a cloud in which the medical images are stored. The invention is not limited in this regard.
With continuing reference to fig. 5, it can be seen from fig. 5 that the medical image acquiring system provided in this embodiment further includes a communication interface 230 and a communication bus 250, wherein the medical image acquiring device 210, the processor 220, the communication interface 230, and the memory 240 complete communication with each other through the communication bus 250. The communication bus 250 may be a Peripheral Component Interconnect (PCI) bus, an Extended Industry Standard Architecture (EISA) bus, or the like. The communication bus 250 may be divided into an address bus, a data bus, a control bus, and the like. For ease of illustration, only one thick line is shown, but this does not mean that there is only one bus or one type of bus. The communication interface 230 is used for communication between the medical image acquisition system and other devices. In other embodiments, the medical image acquisition device 210 may also communicate with the processor 220, the communication interface 230 and the memory 240 via a wired or wireless network (not shown).
The Processor 220 may be a Central Processing Unit (CPU), other general purpose Processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other Programmable logic device, discrete Gate or transistor logic device, discrete hardware component, etc. The general purpose processor may be a microprocessor or the processor 220 may be any conventional processor or the like, the processor 220 being the control center of the medical image acquisition system and connecting the various parts of the overall medical image acquisition system with various interfaces and lines.
The memory 240 may be used to store the computer program, and the medical image to be delineated and the target segmentation image (target delineation result), and the processor 220 implements various functions of the medical image acquisition system by running or executing the computer program stored in the memory 240 and calling the data stored in the memory 240.
The memory 240 may include non-volatile and/or volatile memory 240. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically Programmable ROM (EPROM), electrically Erasable Programmable ROM (EEPROM), or flash memory. Volatile memory can include Random Access Memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms such as Static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double Data Rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous Link DRAM (SLDRAM), rambus (Rambus) direct RAM (RDRAM), direct Rambus Dynamic RAM (DRDRAM), and Rambus Dynamic RAM (RDRAM), among others.
Code for a computer program to carry out operations for aspects of the present invention may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, smalltalk, C + +, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet service provider).
In summary, compared with the prior art, the method, the device and the system for automatically delineating the target area of the medical image provided by the invention have the following advantages: the invention provides a target area automatic delineation method of a medical image, which locates a target area of the medical image to be delineated by adopting a trained target area locating model so as to obtain multi-channel information (including layer information and morphological information of the target area) of the target area. Therefore, the target area positioning model (such as a convolutional neural network model or a classification network model) with stronger edge recognition capability is adopted, so that the recognition capability of target areas in different forms, such as the recognition capability of target areas with larger layer-to-layer deformation and complex boundary morphological structures, can be effectively improved, and good foundation can be laid for subsequent rapid and accurate image segmentation so as to obtain high-quality target area delineation results through the morphological information of the target area. Furthermore, according to the multi-channel information, the target region segmentation model (such as a convolutional neural network model) is adopted to segment the medical image to be delineated, so as to obtain a target region delineation result of the medical image to be delineated, and because the target region segmentation model can be based on the constraint of morphological information (such as shape information in the morphological information, for example, one of a square and a rectangle) of the target region, the defects that the target region segmentation model is not good at identifying fuzzy boundaries and images with large morphological differences are overcome, so that the target region (such as breast cancer, pelvic tumor and the like) with large morphological differences can be positioned and segmented more quickly and accurately. In summary, according to the method for automatically delineating the target area of the medical image provided by the invention, the target area positioning model is used for positioning the target area and the target area segmentation model is used for segmenting the target area, so that the advantages of the target area positioning model and the target area segmentation model are complemented and organically combined, the whole delineating process does not need manual participation, an end-to-end process is realized, and the delineating efficiency is high and the universality is strong; moreover, the problem of differentiation possibly caused by manual delineation can be reduced, and doctors can be better assisted to improve the quality and efficiency of radiotherapy plan design.
The device and the system for automatically delineating the target area of the medical image provided by the invention belong to the same inventive concept as the method for automatically delineating the target area of the medical image provided by the invention, so that the device and the system at least have all the advantages of the invention, and are not described in detail herein.
It should be noted that the apparatuses and methods disclosed in the embodiments herein can be implemented in other ways. The apparatus embodiments described above are merely illustrative, and for example, the flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods and computer program products according to various embodiments herein. In this regard, each block in the flowchart or block diagrams may represent a module, program or a portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
In addition, the functional modules in the embodiments herein may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.
The above description is only for the purpose of describing the preferred embodiments of the present invention, and is not intended to limit the scope of the present invention, and any variations and modifications made by those skilled in the art based on the above disclosure are within the scope of the present invention. It will be apparent to those skilled in the art that various changes and modifications may be made in the present invention without departing from the spirit and scope of the invention. Thus, if such modifications and variations fall within the scope of the present invention and its equivalent technology, it is intended that the present invention also include such modifications and variations.
Claims (10)
1. A target area automatic delineation method of a medical image is characterized by comprising the following steps:
positioning a target area of a medical image to be sketched by adopting a trained target area positioning model so as to acquire multi-channel information of the target area, wherein the multi-channel information comprises morphological information and layer information corresponding to the morphological information;
and according to the multi-channel information, segmenting the medical image to be sketched by adopting a trained target region segmentation model to obtain a target region sketching result of the medical image to be sketched.
2. The method according to claim 1, wherein the positioning a target region of a medical image to be delineated by using the trained target region positioning model to obtain multi-channel information of the target region comprises:
acquiring morphological distribution of the target area by adopting the target area positioning model according to the slice image of the medical image to be sketched;
obtaining multi-channel information of the target area according to the morphological distribution of the target area; the multichannel information comprises a plurality of channel information, and each channel information comprises morphological information and layer information corresponding to the morphological information.
3. The method for automatic delineation of a target region of claim 1 further comprising: and carrying out post-processing on the initial target area image obtained by segmenting the medical image to be sketched, and taking the post-processed initial target area image as the sketching result of the target area.
4. The automatic target delineation method of any one of claims 1-3 wherein the target localization model is a neural network model; the method comprises the following steps of before positioning a target area of a medical image to be delineated by adopting a trained target area positioning model, training by adopting the following steps to obtain the target area positioning model:
acquiring a first sample set, and dividing the first sample set into a first training sample set and a first testing sample set; wherein each sample in the first set of samples comprises a first original medical image and a positioning label corresponding to a delineation target region of the first original medical image;
taking a pre-built neural network model as an initial target area positioning model;
training the initial target area positioning model based on the first training sample set until a first preset training end condition is met to obtain a candidate target area positioning model;
testing the candidate target area positioning model based on the first test sample set, and judging whether the test result meets a second preset training end condition: if so, taking the candidate target area positioning model as the trained target area positioning model;
if not, adjusting model parameters of the candidate target area positioning model, and taking the candidate target area positioning model after parameter adjustment as the initial target area positioning model;
and repeating the steps of training, testing and adjusting the model parameters of the candidate target area positioning model until the test result meets the second preset training end condition to obtain the trained target area positioning model.
5. The method according to claim 4, wherein the localization label of the delineated target area of each sample in the first sample set is obtained by:
acquiring the first original medical image and a delineation target area of the first original medical image;
acquiring the position of the layer surface with the morphological difference exceeding a preset threshold value according to the delineation target area;
and dividing the first original medical image according to the layer position of which the morphological difference exceeds a preset threshold value to obtain the morphological distribution of the delineation target area so as to serve as a positioning label of the delineation target area of the first original medical image.
6. The method of claim 4, wherein the target segmentation model comprises a convolutional neural network model; before the trained target segmentation model is adopted to segment the medical image to be delineated, the method further comprises the following steps of training to obtain the target segmentation model:
acquiring a second sample set, and dividing the second sample set into a second training sample set and a second testing sample set; wherein each sample in the second sample set comprises a second original medical image, a target label of the second original medical image, and multi-channel information corresponding to the target label;
taking a pre-built convolutional neural network model as an initial target area segmentation model;
training the initial target segmentation model based on the second training sample set until a third preset training end condition is met to obtain a candidate target segmentation model;
testing the candidate target segmentation model based on the second test sample set, and judging whether a test result meets a fourth preset training end condition, if so, taking the candidate target segmentation model as the trained target segmentation model;
if not, adjusting the model parameters of the candidate target area segmentation model, and taking the candidate target area segmentation model after the parameters are adjusted as the initial target area segmentation model;
and repeating the steps of training, testing and adjusting the model parameters of the candidate target area segmentation model until the test result meets the fourth preset training end condition to obtain the trained target area segmentation model.
7. The method according to claim 6, wherein before the pre-constructed convolutional neural network model is used as the initial target segmentation model, the method further comprises:
calculating the number of classification categories of morphological information of the target area delineated by any sample in the first sample set or calculating the number of classification categories of morphological information of multi-channel information of the target area label of any sample in the second sample set;
taking the number of the classification categories as the number of channels of the convolutional neural network model;
and building and initializing the convolutional neural network model according to the number of channels of the convolutional neural network model.
8. The method for automatic delineation of a target region of claim 1 further comprising:
performing first pretreatment on each sample in the obtained first sample set to obtain the target area positioning model by using the first sample set after the first pretreatment;
and/or performing second preprocessing on each sample in the acquired second sample set to train the target region segmentation model by using the second preprocessed second sample set.
9. An automatic target delineation device of medical image, comprising:
the target area positioning module is configured to position a target area of a medical image to be sketched by adopting a trained target area positioning model so as to acquire multi-channel information of the target area, wherein the multi-channel information comprises morphological information and layer information corresponding to the morphological information;
and the target area segmentation module is configured to segment the medical image to be sketched by adopting a trained target area segmentation model according to the multi-channel information to obtain a target area sketching result of the medical image to be sketched.
10. A medical image acquisition system comprising a medical image acquisition device, a processor and a memory, the memory having stored thereon a computer program;
the medical image acquisition device is configured to acquire a medical image to be sketched; when the computer program is executed by the processor, the target region automatic delineation method of the medical image according to any one of claims 1 to 8 is adopted to carry out target region delineation on the medical image to be delineated, and a target region delineation result of the medical image to be delineated is obtained.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211282046.8A CN115762724A (en) | 2022-10-19 | 2022-10-19 | Method, device and system for automatically delineating target area of medical image |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211282046.8A CN115762724A (en) | 2022-10-19 | 2022-10-19 | Method, device and system for automatically delineating target area of medical image |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115762724A true CN115762724A (en) | 2023-03-07 |
Family
ID=85353937
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211282046.8A Pending CN115762724A (en) | 2022-10-19 | 2022-10-19 | Method, device and system for automatically delineating target area of medical image |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115762724A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117351489A (en) * | 2023-12-06 | 2024-01-05 | 四川省肿瘤医院 | Head and neck tumor target area delineating system for whole-body PET/CT scanning |
-
2022
- 2022-10-19 CN CN202211282046.8A patent/CN115762724A/en active Pending
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117351489A (en) * | 2023-12-06 | 2024-01-05 | 四川省肿瘤医院 | Head and neck tumor target area delineating system for whole-body PET/CT scanning |
CN117351489B (en) * | 2023-12-06 | 2024-03-08 | 四川省肿瘤医院 | Head and neck tumor target area delineating system for whole-body PET/CT scanning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP6567179B2 (en) | Pseudo CT generation from MR data using feature regression model | |
US11062449B2 (en) | Method and system for extracting vasculature | |
CN109493951B (en) | System and method for reducing radiation dose | |
CN111008984B (en) | Automatic contour line drawing method for normal organ in medical image | |
JP7030050B2 (en) | Pseudo-CT generation from MR data using tissue parameter estimation | |
US10229517B2 (en) | Method and apparatus for automated determination of contours in iterative reconstruction of image data | |
WO2021213519A1 (en) | Image obtaining method and system, image quality determination method and system, and medical image acquisition method and system | |
John et al. | Multilevel thresholding based segmentation and feature extraction for pulmonary nodule detection | |
US20220092787A1 (en) | Systems and methods for processing x-ray images | |
CN110739049A (en) | Image sketching method and device, storage medium and computer equipment | |
CN112601582A (en) | System and method for accelerated on-line adaptive radiotherapy | |
CN115666717A (en) | Method and apparatus for deriving and using virtual volume structures to predict potential collisions in delivering therapeutic radiation | |
CN115762724A (en) | Method, device and system for automatically delineating target area of medical image | |
Lappas et al. | Automatic contouring of normal tissues with deep learning for preclinical radiation studies | |
US11534623B2 (en) | Determining at least one final two-dimensional image for visualizing an object of interest in a three dimensional ultrasound volume | |
CN110599530A (en) | MVCT image texture enhancement method based on double regular constraints | |
US20080285822A1 (en) | Automated Stool Removal Method For Medical Imaging | |
CN118302791A (en) | Machine learning based segmentation of biological objects in medical images | |
Wodzinski et al. | Application of B-splines FFD image registration in breast cancer radiotherapy planning | |
Rodriguez Ruiz | Artificial intelligence and tomosynthesis for breast cancer detection | |
CN117427286B (en) | Tumor radiotherapy target area identification method, system and equipment based on energy spectrum CT | |
大矢めぐみ et al. | Investigation of clinical target volume segmentation for whole breast irradiation using three-dimensional convolutional neural networks with gradient-weighted class activation mapping | |
Ramalingam et al. | Estimating and locating the volumes of multiple primary non-small cell stage-I bronchogenic carcinomas from medical images using specially devised morphological operations |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |