CN114359317A - Blood vessel reconstruction method based on small sample identification - Google Patents
Blood vessel reconstruction method based on small sample identification Download PDFInfo
- Publication number
- CN114359317A CN114359317A CN202111550766.3A CN202111550766A CN114359317A CN 114359317 A CN114359317 A CN 114359317A CN 202111550766 A CN202111550766 A CN 202111550766A CN 114359317 A CN114359317 A CN 114359317A
- Authority
- CN
- China
- Prior art keywords
- blood vessel
- data
- vessel
- identification
- slice
- 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
- Image Analysis (AREA)
- Apparatus For Radiation Diagnosis (AREA)
Abstract
The invention discloses a blood vessel reconstruction method based on small sample identification, which comprises the following steps: s1, acquiring CT data of a training sample vessel tree, wherein the CT data comprises CT slices with vessel cross-section images and corresponding CT values; s2, labeling the blood vessel in the partial CT slice on the equipment by the operator to obtain a group of fine labeling data; s3, based on the fine labeling data obtained in S2, a primary recognition model is obtained through neural network training; s4, iterative data expansion training is carried out on the primary recognition model, and finally a recognition model capable of recognizing blood vessels in the CT slice and outputting a result is obtained; s5, based on the identification model obtained in S4, identifying the CT data of the blood vessel tree to be reconstructed to obtain the spatial information of the blood vessel; and S6, connecting the blood vessels according to the spatial information obtained in S5, and obtaining a final three-dimensional model of the blood vessel tree through post processing. The method reduces the time consumed by manual labeling while ensuring the identification accuracy, and improves the identification efficiency.
Description
Technical Field
The invention relates to the technical field of medical image processing, in particular to a blood vessel reconstruction method based on small sample identification.
Background
The identification, the segmentation and the reconstruction of the pulmonary artery and vein vessels play a crucial role in the minimally invasive puncture operation applied to the pulmonary tumor treatment, when a rigid needle body of the puncture operation enters the body of a patient, the puncture needle mistakenly touches the artery and the vein to cause large bleeding in the operation due to the posture change caused by the respiration of the patient in the operation, and therefore accurate and clear intra-operation vessel image guidance is necessary.
At present, tasks such as blood vessel segmentation and the like focus on parts such as eyeballs, hearts, livers and the like, while the prior art uses a blood vessel tree as a training sample to carry out end-to-end network learning, the method has the problems of large requirements on training data quantity, long training time, high dependence on the sample and the network and the like.
Patent document CN113674279A discloses a method and apparatus for processing coronary CTA images based on deep learning, the method includes converting a CTA image sequence into a target NIFTI file for identification, and converting the obtained NIFTI file with mask information into a target mask image sequence; removing a sternum area in the CTA image sequence according to the target mask image sequence to obtain a target image sequence; carrying out image three-dimensional reconstruction based on volume rendering on the target image sequence; and extracting a blood vessel region from the three-dimensional model, and projecting each point of the blood vessel region onto a two-dimensional plane to obtain a reconstructed image after the blood vessel is straightened. According to the method, the spatial three-dimensional coordinates of the blood vessel are extracted from the whole three-dimensional modeling, then modeling is carried out based on the three-dimensional coordinates, the early preparation work takes too long, and certain requirements are met for the operational capability of equipment.
Patent document CN109063557B discloses a method for quickly constructing identification data of coronary artery blood vessels of heart, which includes acquiring an original picture; performing pixel-level labeling on a very small number of original pictures according to the rough label pictures to form fine label pictures; changing the fine labeling picture from a three-channel image into a single-channel image; carrying out binarization processing on the single-channel image and storing the single-channel image into a binarization picture; taking the corresponding original picture of the binary picture as training data, training the initial network until the initial network is reached; inputting all original pictures into the first network to obtain a binarization result picture, and generating a pseudo fine labeling picture based on the binarization result picture and a corresponding coarse labeling picture; and establishing a database based on the original picture, the fine labeling picture and the pseudo labeling picture so as to train the network. According to the method, a pseudo-label graph is built to generate a network to replace manual labeling, but the network can only identify the blood vessels in the 2D image, cannot construct a 3D blood vessel tree, and cannot provide accurate and reliable spatial information for puncture.
Disclosure of Invention
In order to solve the problems, the invention provides a blood vessel reconstruction method based on small sample recognition, which only needs to manually mark a small amount of precise marking data as an initial training material, subsequently uses a result output by a recognition model as alternative data, and uses the alternative data as a secondary training material after manual correction, thereby ensuring the recognition accuracy, reducing the time consumption of manual marking and improving the recognition efficiency.
A method of vessel reconstruction based on small sample identification, comprising:
s1, acquiring CT data of a training sample vessel tree, wherein the CT data comprises CT slices with vessel cross-section images and corresponding CT values;
s2, labeling the blood vessel in the partial CT slice on the equipment by the operator to obtain a group of fine labeling data;
s3, based on the fine labeling data obtained in S2, a primary recognition model is obtained through neural network training;
s4, iterative data expansion training is carried out on the primary recognition model, and finally a recognition model capable of recognizing blood vessels in the CT slice and outputting a result is obtained;
s5, based on the identification model obtained in S4, identifying the CT data of the blood vessel tree to be reconstructed to obtain the spatial information of the blood vessel;
and S6, connecting the blood vessels according to the spatial information obtained in S5, and obtaining a final three-dimensional model of the blood vessel tree through post processing.
Preferably, the iterative data expansion training in S4 includes:
s4.1, selecting some unmarked CT slices, putting the slices into a primary identification model for identification, and obtaining alternative data consisting of identified blood vessel coordinate points;
s4.2, an operator carries out manual review on the alternative data on the equipment to obtain a group of pseudo-fine labeled data;
s4.3, putting the pseudo fine labeling data obtained in the S4.2 into a primary recognition model for training;
and S4.4, repeating the steps until the preset identification accuracy is reached, and obtaining a final identification model, wherein the preset identification accuracy is equal to or more than 80%.
Preferably, the labeling is to add a recognition mark to a central point of a cross section of the blood vessel in the CT slice and record coordinates thereof.
Preferably, the manual rechecking in S4.2 is specifically that an operator determines a coordinate point in the candidate data, wherein the correctly identified coordinate point is used as a positive sample, the incorrectly identified coordinate point is used as a negative sample, and the positive sample and the negative sample are integrated to be pseudo-fine labeling data; because the initial accuracy of the recognition model trained by small sample recognition is not high, the positive sample proportion is not high in the output result of the initial recognition model, so that an operator only needs to add a new recognition identifier to the coordinate point of the positive sample when rechecking the alternative data, and thus pseudo-fine labeling data with high accuracy is obtained; furthermore, the pseudo-fine labeling data with high accuracy is used as a secondary training sample, so that the training effect of the recognition model is better.
Preferably, the S1 normalizes the gray value of the CT slice; the blood vessel image is more obvious through the adjustment of the pixels, and the identification is convenient.
Preferably, the output result in S4 includes a CT slice with an identifier, where the identifier is a two-dimensional coordinate of the CT slice and a CT value corresponding to the coordinate point.
Preferably, the spatial information in S5 includes two-dimensional coordinates of the center point of the blood vessel and the corresponding serial number of the CT slice, wherein the serial number is taken as a vertical coordinate.
Preferably, the specific process of S6 is:
s6.1, by setting the CT value of each blood vessel central point as a standard set by a self-adaptive threshold, extending and searching for a difference coordinate point which is obviously different from the CT value of the blood vessel central point to the periphery;
s6.2, forming a closed rectangle based on the distinguishing coordinate points, continuously identifying all blood vessel points in the closed rectangle, and finally obtaining a blood vessel contour;
s6.3, connecting the blood vessel contours adjacent to the serial number of the CT slice according to a near principle and a gradient consistency principle to obtain a blood vessel model;
and S6.4, integrating all the blood vessel models and outputting the integrated blood vessel models to obtain a three-dimensional model of the blood vessel tree.
Compared with the prior art, the invention has the beneficial effects that:
(1) the identification model is subjected to closed-loop training based on the small sample blood vessel two-dimensional image, so that the accuracy of the identification model can be continuously improved.
(2) In the sample marking process, only a small amount of manual accurate marking data are used as training samples, and subsequently, only correct coordinate points in the identification result need to be judged, and re-marking is carried out to be used as secondary training samples, so that the time consumption of manual marking is reduced, and the accuracy of the training samples is guaranteed.
(3) The recognition model is used for recognizing and generating three-dimensional data based on the two-dimensional image, is different from the prior art for recognizing the direct three-dimensional image, and has smaller operation pressure on equipment.
Drawings
FIG. 1 is a schematic flow diagram of a method for reconstructing a blood vessel provided by the present invention;
FIG. 2 is an effect diagram of a three-dimensional reconstruction model of a pulmonary vascular tree generated by the technical solution of the present invention.
Detailed Description
As shown in fig. 1, a blood vessel reconstruction method based on small sample identification specifically includes the steps of:
s1 CT data of a training sample vessel tree are obtained, wherein the CT data comprise CT slices with vessel cross-section images and corresponding CT values, window width and window level adjustment is carried out on the CT slices before training, vessel development is obvious, gray values are normalized, and the sizes of the images are unified to be proper resolution.
Wherein window width window level refers to, for example, a density difference that CT can identify 2000 different gray levels in a human body. However, the human eye can only distinguish 16 gray scales, that is, the CT value that the human eye can distinguish on the CT image should be 125Hu (2000/16), that is, the CT values of different tissues in the human body can only be recognized by the human eye if the difference is more than 125 Hu. The CT value of human soft tissue is changed between 20-50Hu, so human eyes can not recognize the change, and the advantage of CT can be reflected only by sectional observation.
The observed CT value range is called window width, and the observed central CT value is the window level or window center.
Window width window level adjustment may be operationally equivalent to setting the lower threshold to-150, the gray scale values less than-150 to-150, the upper threshold to 200, and the gray scale values greater than 200 to 200.
After the normalization processing statistics is carried out to obtain the maximum and minimum values of all pixels after the window width and the window level are adjusted, min-max standardization is carried out on each pixel, and the result value is mapped to the range from [0 to 1], wherein the min-max standardization transfer function is as follows:
wherein max is the maximum value of the sample data, min is the minimum value of the sample data, x is the pixel value before normalization processing, and y is the pixel value after normalization processing.
S2, labeling the blood vessels in a certain random number of CT slices on equipment by an operator to obtain a group of fine-labeled data, wherein the operator only needs to add an identification mark to the central point of the cross section of the blood vessel in the CT slice;
s3 obtains an initial recognition model through neural network training based on the fine-scale data obtained in S2, where the initial recognition model also has the ability to recognize blood vessels in any slice in the CT data, but the accuracy of the initial recognition model is not high.
S4, iterative data expansion training is carried out on the initial blood vessel recognition model, and finally the recognition model capable of recognizing blood vessels in the CT slice and outputting results is obtained, wherein the iterative data expansion training specifically comprises the following steps:
s4.1, selecting some unmarked CT slices, putting the slices into a primary identification model for identification, and obtaining alternative data consisting of identified blood vessel coordinate points;
s4.2, the operator manually rechecks the alternative data on the equipment:
judging coordinate points in the alternative data by an operator, wherein the coordinate points which are correctly identified are used as positive samples, the coordinate points which are incorrectly identified are used as negative samples, and marking new identification marks on the coordinate points corresponding to the positive samples to obtain a group of pseudo-fine marking data;
s4.3, putting the pseudo fine labeling data obtained in the S4.2 into a primary recognition model for training;
and S4.4, repeating the steps until the preset identification accuracy is reached to obtain a final identification model, wherein the iteration is stopped when the preset identification accuracy is more than 80%.
The output result of the recognition model comprises a CT slice with a mark, a two-dimensional coordinate of a central point of a blood vessel corresponding to the mark on the CT slice and a CT value corresponding to the central point of the blood vessel.
And S5, based on the identification model obtained in S4, identifying the CT data of the blood vessel tree to be reconstructed, and obtaining the two-dimensional coordinates of the center point of the blood vessel and the corresponding serial number of the CT slice, wherein the serial number is used as the vertical coordinate.
S6, based on the two-dimensional coordinates of the blood vessel center point obtained in S5 and the corresponding CT slice serial number, three-dimensional reconstruction of the blood vessel tree is carried out:
s6.1, by setting the CT value of each blood vessel central point as a standard set by a self-adaptive threshold, extending and searching for a difference coordinate point which is obviously different from the CT value of the blood vessel central point to the periphery;
s6.2, forming a closed rectangle based on the distinguishing coordinate points, continuously identifying all blood vessel points in the closed rectangle, and finally obtaining a blood vessel contour;
s6.3, connecting the blood vessel contours adjacent to the serial number of the CT slice according to a near principle and a gradient consistency principle to obtain a blood vessel model;
s6.4, integrating all the blood vessel models and then outputting the integrated blood vessel models to obtain a three-dimensional model of the blood vessel tree, wherein the actual modeling effect of the three-dimensional model is shown in figure 2.
Claims (8)
1. A blood vessel reconstruction method based on small sample identification is characterized by comprising the following steps:
s1, acquiring CT data of a training sample vessel tree, wherein the CT data comprises CT slices with vessel cross-section images and corresponding CT values;
s2, labeling the blood vessel in the partial CT slice on the equipment by the operator to obtain a group of fine labeling data;
s3, based on the fine labeling data obtained in S2, a primary recognition model is obtained through neural network training;
s4, iterative data expansion training is carried out on the primary recognition model, and finally a recognition model capable of recognizing blood vessels in the CT slice and outputting a result is obtained;
s5, based on the identification model obtained in S4, identifying the CT data of the blood vessel tree to be reconstructed to obtain the spatial information of the blood vessel;
and S6, connecting the blood vessels according to the spatial information obtained in S5, and obtaining a final three-dimensional model of the blood vessel tree through post processing.
2. The vessel reconstruction method according to claim 1, wherein the iterative data augmentation training in S4 is specifically:
s4.1, selecting some unmarked CT slices, putting the slices into a primary identification model for identification, and obtaining alternative data consisting of identified blood vessel coordinate points;
s4.2, an operator carries out manual review on the alternative data on the equipment to obtain a group of pseudo-fine labeled data;
s4.3, putting the pseudo fine labeling data obtained in the S4.2 into a primary recognition model for training;
and S4.4, repeating the steps until the preset identification accuracy is reached, and obtaining the final identification model.
3. The vessel reconstruction method according to claim 1 or 2, wherein the labeling is to add a recognition mark to the center point of the vessel cross section in the CT slice and record the coordinates thereof.
4. The vessel reconstruction method according to claim 2, wherein in S4.2, the manual review is performed, specifically, an operator determines coordinate points in the candidate data, and adds a new identification mark to the correctly identified vessel coordinate points.
5. The vessel reconstructing method according to claim 1, wherein said S1 is configured to normalize the gray-level values of the CT slices.
6. The vessel reconstructing method according to claim 1, wherein the output result in S4 includes a CT slice with an identifier, the identifier is a two-dimensional coordinate of the CT slice and a CT value corresponding to the coordinate point.
7. The vessel reconstructing method according to claim 1, wherein the spatial information in S5 includes two-dimensional coordinates of a center point of the vessel, and a serial number of a corresponding CT slice, wherein the serial number is taken as a vertical coordinate.
8. The vessel reconstruction method according to claim 1, wherein the S6 specific process:
s6.1, by setting the CT value of each blood vessel central point as a standard set by a self-adaptive threshold, extending and searching for a difference coordinate point which is obviously different from the CT value of the blood vessel central point to the periphery;
s6.2, forming a closed rectangle based on the distinguishing coordinate points, continuously identifying all blood vessel points in the closed rectangle, and finally obtaining a blood vessel contour;
s6.3, connecting the blood vessel contours adjacent to the serial number of the CT slice according to a near principle and a gradient consistency principle to obtain a blood vessel model;
and S6.4, integrating all the blood vessel models and outputting the integrated blood vessel models to obtain a three-dimensional model of the blood vessel tree.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111550766.3A CN114359317A (en) | 2021-12-17 | 2021-12-17 | Blood vessel reconstruction method based on small sample identification |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111550766.3A CN114359317A (en) | 2021-12-17 | 2021-12-17 | Blood vessel reconstruction method based on small sample identification |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114359317A true CN114359317A (en) | 2022-04-15 |
Family
ID=81099940
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111550766.3A Pending CN114359317A (en) | 2021-12-17 | 2021-12-17 | Blood vessel reconstruction method based on small sample identification |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114359317A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117079080A (en) * | 2023-10-11 | 2023-11-17 | 青岛美迪康数字工程有限公司 | Training optimization method, device and equipment for coronary artery CTA intelligent segmentation model |
-
2021
- 2021-12-17 CN CN202111550766.3A patent/CN114359317A/en active Pending
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117079080A (en) * | 2023-10-11 | 2023-11-17 | 青岛美迪康数字工程有限公司 | Training optimization method, device and equipment for coronary artery CTA intelligent segmentation model |
CN117079080B (en) * | 2023-10-11 | 2024-01-30 | 青岛美迪康数字工程有限公司 | Training optimization method, device and equipment for coronary artery CTA intelligent segmentation model |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN102890823B (en) | Motion object outline is extracted and left ventricle image partition method and device | |
CN106485695B (en) | Medical image Graph Cut dividing method based on statistical shape model | |
CN109753997B (en) | Automatic accurate robust segmentation method for liver tumor in CT image | |
CN104794708A (en) | Atherosclerosis plaque composition dividing method based on multi-feature learning | |
Mountney et al. | Soft tissue tracking for minimally invasive surgery: Learning local deformation online | |
CN108062749B (en) | Identification method and device for levator ani fissure hole and electronic equipment | |
US9730609B2 (en) | Method and system for aortic valve calcification evaluation | |
CN106846346A (en) | Sequence C T image pelvis profile rapid extracting methods based on key frame marker | |
KR101223681B1 (en) | Automatic Segmentation device and method of Cartilage in Magnetic Resonance Image | |
CN112184720B (en) | Method and system for segmenting internal rectus muscle and optic nerve of CT image | |
WO2017086433A1 (en) | Medical image processing method, device, system, and program | |
CN110334566A (en) | Fingerprint extraction method inside and outside a kind of OCT based on three-dimensional full convolutional neural networks | |
CN106780491A (en) | The initial profile generation method used in GVF methods segmentation CT pelvis images | |
CN114359317A (en) | Blood vessel reconstruction method based on small sample identification | |
CN106780492B (en) | Method for extracting key frame of CT pelvic image | |
CN115578320A (en) | Full-automatic space registration method and system for orthopedic surgery robot | |
CN108898601B (en) | Femoral head image segmentation device and method based on random forest | |
CN116309647B (en) | Method for constructing craniocerebral lesion image segmentation model, image segmentation method and device | |
CN105225234A (en) | Based on the lung tumor identification method of support vector machine MRI Iamge Segmentation | |
CN110910409A (en) | Gray scale image processing method and device and computer readable storage medium | |
CN110610502A (en) | Automatic aortic arch region positioning and segmentation method based on CT image | |
Krawczyk et al. | YOLO and morphing-based method for 3D individualised bone model creation | |
CN115115662A (en) | Knowledge and data fusion drive-based multi-modal medical image target segmentation method | |
CN111145353B (en) | Method for generating 3D point cloud through image segmentation and grid feature point extraction algorithm | |
Wang et al. | A machine learning approach to extract spinal column centerline from three-dimensional CT data |
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 |