CN112767374A

CN112767374A - Alzheimer disease focus region semantic segmentation algorithm based on MRI

Info

Publication number: CN112767374A
Application number: CN202110108261.5A
Authority: CN
Inventors: 赵萌; 敖吉; 杨虹霞; 石凡; 陈胜勇
Original assignee: Tianjin University of Technology
Current assignee: Tianjin University of Technology
Priority date: 2021-01-27
Filing date: 2021-01-27
Publication date: 2021-05-07

Abstract

An Alzheimer disease focus region semantic segmentation algorithm based on MRI belongs to the field of computer vision and medical image processing. Based on the deep convolutional neural network, a brain MRI data set is newly established, various focus regions are labeled, a network model shown in figure 2 is trained, and the problem of multi-scale focus target segmentation is further solved on the basis of completing a basic semantic segmentation task. As shown in fig. 2, the network model takes a general semantic segmentation network as a basic framework, enhances feature expression and attention to salient features, and performs feature extraction on lesion regions of different scales to more accurately determine the position of each lesion region. The method is suitable for multi-modal 3D medical data, such as MRI or CT images of organs such as brain, heart and the like, and is used for semantic segmentation of focus regions.

Description

Alzheimer disease focus region semantic segmentation algorithm based on MRI

Technical Field

The invention relates to an MRI-based Alzheimer disease focus region semantic segmentation algorithm, and belongs to the field of computer vision and medical image processing.

Background

Magnetic Resonance Imaging (MRI) is a well-known method in radiology widely used for visualizing internal structures of the human body. MRI uses magnets, radio waves, and computer technology to produce images. One of the advantages of magnetic resonance over other imaging techniques is that it does not use ionizing radiation, so that it provides relatively good contrast between different soft tissues of the body. But the image quality is poor, so that great challenges are brought to accurate image segmentation. Therefore, some pre-treatment is usually performed to improve the results of the subsequent steps. The purpose of segmentation in medical image processing is to select some organs, such as the liver, parts of joints or the brain. This work involves brain segmentation and volume measurements (volumetric method).

Quantitative analysis of brain images is a routine examination of many neurological diseases and conditions. Segmentation, i.e. labeling 2D pixels (or 3D pixels), is a key component of quantitative analysis. Manual segmentation requires a layer-by-layer analysis of its structure, which is expensive and cumbersome, and also presents some human errors. Therefore, an automatic segmentation method is needed to achieve higher accuracy.

As 3D and 4D imaging become more conventional, and as physiological and functional imaging increases, the size and complexity of medical imaging data also increases. Therefore, it is critical to develop tools that can help extract information from these large datasets. Machine learning is a set of algorithmic techniques that allow computer systems to perform data-driven predictions from large data. These techniques have a variety of applications that can be tailored to the medical field.

Computer-aided diagnosis has been a valuable automated diagnostic tool, and the study of artificial intelligence technology represented by deep learning combined with medical big data has been successful in the biomedical field. The classification and segmentation algorithm based on the 3D convolutional neural network can effectively classify patients with different degrees, such as mild cognitive impairment, dementia, Alzheimer's Disease and the like, can predict a lesion area, and can assist doctors in clinical diagnosis by predicting early Alzheimer's Disease (AD) with the assistance of a computer.

AD, a common degenerative disease of the central nervous system of the brain, has an increasing incidence with age. The incubation period of AD is long, and its clinical symptoms gradually worsen with time, including memory loss, cognitive impairment, dementia, and the like. At present, the mainstream diagnosis method of AD depends on comprehensive analysis and judgment of clinical data by physicians, and although these methods all achieve good diagnosis effects, they are time-consuming and labor-consuming after all, and have certain subjectivity, so that misdiagnosis still may occur. With current medical approaches, AD is not yet cured, but if it can be diagnosed correctly, the correct treatment can be used to delay the patient's condition, so early diagnosis and prognosis of AD is a critical tool in controlling the disease.

Some existing medical assessment techniques for AD include physical and neurobiological examination, Mini Mental State Exam (MMSE). The use of brain MRI images to detect AD has attracted more attention due to the possibility of early treatment or intervention. In order to detect AD patients, the feature learning/extraction method can be roughly divided into two cases according to the way of extracting features by a manual design technique and by an automatic method, in addition to other clinical symptoms. The former is typically associated with analytical or model-based methods, while the latter is associated with deep learning methods. Since the manual functions are largely dependent on the knowledge of human experts, the automatic learning of functions through a deep learning method has attracted great interest.

Disclosure of Invention

The invention aims to provide an Alzheimer disease focus region semantic segmentation algorithm based on MRI (magnetic resonance imaging) so as to realize detection and classification of six focus regions in an MRI brain image.

In order to achieve the purpose, the scheme of the invention is as follows:

a semantic segmentation algorithm for a focus region of Alzheimer's disease based on MRI takes a general semantic segmentation network as a basic frame, strengthens feature expression and attention to salient features, and extracts features aiming at focus regions with different scales so as to more accurately determine the position of each focus region. The current concrete steps are as follows:

(1) the method comprises the steps that a hospital doctor finishes collection of brain MRI image samples of clinical patients, the brain MRI image samples are divided into two sequences of T1 and T2, images in a dcm format are derived through professional software, each dcm image is a 2D image, the resolution is 512 x 512, 20 dcm images of each patient are collected, and MRI data of 100 patients are collected;

(2) the labeling of the focus area of each dcm image is completed by a pathology specialist, and the focus areas are divided into six types: labeling frontal lobe, temporal lobe, parietal lobe, hippocampus, midbrain and hemioval center with red, green, blue, yellow, cyan and purple respectively, and exporting and storing with 3D Slicer4.10.2 and ITK-SNAP software;

(3) manually combining 20 2D dcm images of each patient into one 3D nii format image, i.e. the MRI data for each patient is nii images with a resolution of 512 x 20;

(4) a neural network is designed, so that a basic semantic segmentation function can be realized;

(5) based on the characteristics of a data set, the problem of a focus area with various scales is considered, and a special feature extraction network is designed;

(6) multiple experiments are performed, the optimal network parameters are explored, and various evaluation index evaluation models are used, such as: precision (Precision), Recall (Recall), similar entropy (Dice), etc.;

(7) designing an ablation experiment, and exploring the specific problem solved by each module;

(8) the feasibility and the superiority of the method are proved, namely the method can realize the correct detection and classification of all lesion areas in the MRI image.

The invention has the beneficial effects that: the method can effectively realize semantic segmentation of all focus areas in the MRI image. Based on a neural network, a brain MRI data set is newly established, various focus areas are labeled, a suitable network model is trained, and a challenging problem is further solved on the basis of completing a basic semantic segmentation task: and (5) multi-scale lesion target segmentation. In order to solve the problem, a new feature extraction and fusion module is constructed, feature expression is enhanced, different feature channels are weighted, and channels containing more effective information have larger weights. When the characteristics are fused, more attention is paid to the focus region under a specific scale, so that the detection of the multi-scale focus region is realized. The method is feasible and can be applied to the field of semantic segmentation of MRI images of other organs through fine adjustment.

Brief description of the drawings

FIG. 1 is a flow chart of the production of a data set.

FIG. 2 is a flow chart of the MRI-based Alzheimer disease lesion region semantic segmentation algorithm of the present invention.

Detailed Description

A semantic segmentation algorithm of a focus region of Alzheimer's disease based on MRI, based on the semantic segmentation algorithm of the focus region of brain MRI image, solve the classification and recognition problem of multiple focus regions in the MRI image, do little power for medical research.

Magnetic resonance imaging is a type of tomographic imaging that uses the magnetic resonance phenomenon to acquire electromagnetic signals from a human body and reconstruct human body information. A magnetic resonance imaging technique spatially encodes nuclear magnetic resonance signals, which can reconstruct images of the human body. Magnetic resonance imaging can obtain tomographic images in any direction, three-dimensional volume images, and even four-dimensional images of space-spectrum distribution.

After MRI data was collected from the hospital, each dcm image was derived, for six lesion areas per image: the frontal lobe, temporal lobe, parietal lobe, hippocampus, midbrain, and hemioval center are separately labeled, and then the dcm images are synthesized into nii images as input to the neural network, as shown in fig. 1.

Based on the labeled MRI 3D ni image, semantic segmentation is realized by using a Convolutional Neural Network (CNN) and a deep learning method, and feature learning is performed on various focus areas in the MRI image. The convolutional neural network is a feedforward neural network which comprises convolutional calculation and has a deep structure, and is one of representative algorithms of deep learning.

The web framework employs 3D U-Net, after which a residual mechanism and attention mechanism are introduced during the down-sampling and up-sampling phases. We have 100 MRI 3D ni images as dataset and 90 images as training set, where each 3D ni image contains 20 2D dcm images, each dcm image contains six lesion regions: frontal lobe, temporal lobe, parietal lobe, hippocampus, midbrain, and hemioval center. As shown in fig. 2, T1 (longitudinal duration) mode and T2 (transverse duration) mode samples of brain MRI data of clinical patients are first collected, and modal weighting is performed to train a semantic segmentation neural network. Firstly, the semantic information is restored through feature extraction network and up-sampling, then feature fusion is carried out on the semantic information and the original image, the significant features under different scales are obtained, and then subsequent classification and segmentation tasks are carried out. And calculating Precision (Precision), Recall (Recall) and similar entropy (Dice) according to the detected confusion matrix, and evaluating the performance of the model.

In our neural network algorithm, based on 3D U-Net network framework, 5 layers of down sampling are firstly used for extracting effective features, each layer of operation comprises residual connecting blocks of 1 × 1, 3 × 3 and 5 × 5 with different scales, the layer of output is obtained through a plurality of convolution operations and nonlinear operations, each layer of output is more accurately emphasized through a channel attention mechanism and a space attention mechanism, and interference of background and other noises on effective feature information is avoided.

After effective features are extracted, semantic information of each focus area is restored through 5-layer upper sampling operation, feature fusion is carried out on the semantic information and original images with different scales, and feature graphs under different scales are accumulated layer by layer, so that the network can learn deep global semantic features and can keep shallow local contour information. Finally, semantic segmentation results of the six types of focus regions with different scales are obtained.

And (3) evaluating the detection result of the model by adopting Precision (Precision), Recall (Recall) and similar entropy (Dice) on the basis of the confusion matrix:

wherein, tp (true positive) represents true positive case, fn (false negative) represents false negative case, fp (false positive) represents false positive case, and they represent different prediction results. The accuracy reflects the accuracy of the model, the recall rate reflects the comprehensiveness of the model, the Dice coefficient reflects the similarity of two sets, wherein X represents TP + FP, Y represents TP + FN, and the accuracy and the recall rate are comprehensively considered. The higher the result, the better the effect.

The method comprises the following specific steps:

(1) collecting T1 (longitudinal duration) weighted imaging and T2 (transverse duration) weighted imaging sequence samples of brain MRI data of a clinical patient and deriving dcm-formatted data;

(2) labeling all focus regions in the dcm image by a pathology specialist, wherein the labeling includes focus region categories and accurate edge labeling;

(3) synthesizing the 2D dcm data of each patient into 3D ni data, performing preprocessing operations on all nii images, including cutting to a uniform size to be used as input of a neural network, performing random rotation, stretching and the like on the images, and normalizing and standardizing the images through a preset value;

(4) a convolutional neural network suitable for an existing data set is designed to realize semantic segmentation of focus regions, feature learning is firstly carried out on various focus regions in an MRI image through a feature extraction network, effective focus region feature information is extracted, and then a subsequent segmentation task is completed. The characteristic extraction network can also be other convolutional neural networks, and can also perform characteristic learning on various focus areas in the MRI image;

(5) the network performance is evaluated by calculating Precision, Recall and Dice indexes, and the superiority of the segmentation performance is proved.

It should be noted that the above-mentioned embodiments are only examples of the present invention, and are only illustrative of the present invention, and therefore do not limit the scope of the present invention. The technical idea of the invention is that only obvious changes are needed and still fall within the scope of the invention.

Claims

1. An MRI-based Alzheimer disease lesion region semantic segmentation algorithm is characterized by comprising the following steps of:

(1) collecting T1 and T2 sequence samples of brain MRI data of a clinical patient and deriving dcm-formatted image data;

(2) marking all focus areas in the dcm image, including focus area categories and accurate edge marking thereof;

(3) synthesizing the 2D dcm data of each patient into 3D ni image data, performing preprocessing operations on all nii images, including cutting to a uniform size to be used as input of a neural network, randomly rotating and stretching the images, and normalizing the images by a preset value;

(4) designing a neural network suitable for the existing data set to realize semantic segmentation of a focus area;

(5) and evaluating the performance of the neural network through calculating accuracy, recall rate and similar entropy indexes.

2. The MRI-based semantic segmentation algorithm for the lesion areas of alzheimer's disease according to claim 1, wherein the neural network in step (4) is a feature extraction network to extract effective feature information of the lesion areas, thereby completing subsequent segmentation tasks.

3. The MRI-based lesion region semantic segmentation algorithm of alzheimer's disease as set forth in claim 1, wherein the neural network in the step (4) is a convolutional neural network, and the convolutional neural network performs feature learning on a plurality of lesion regions in the MRI image.

4. The MRI-based semantic segmentation algorithm for the lesion regions of alzheimer's disease as claimed in claim 1, wherein in step (4), based on a 3D U-Net network framework, 5 layers of downsampling are used to extract valid features, each layer of operation includes residual connection blocks of 1 × 1, 3 × 3, 5 × 5 at different scales, the layer of output is obtained through several convolution operations and nonlinear operations, and each layer of output is processed through a channel attention mechanism and a spatial attention mechanism respectively to more accurately emphasize valid features, so as to avoid interference of background and other noise on valid feature information.

5. The MRI-based semantic segmentation algorithm for the lesion areas of Alzheimer's disease according to claim 4, wherein after effective features are extracted, semantic information of each lesion area is restored through 5-layer upsampling operation, feature fusion is performed on the semantic information and original images with different scales obtained in the step (4) during feature extraction, and feature maps with different scales obtained in the network upsampling stage are accumulated layer by layer, so that the network can learn deep global semantic features and can keep shallow local contour information, and finally, semantic segmentation results of six types of lesion areas with different scales are obtained.

6. The MRI-based semantic segmentation algorithm for focal regions of alzheimer's disease as claimed in claim 1, wherein the detection results of the accuracy, recall rate and similarity entropy evaluation model are adopted in step (5) based on the confusion matrix:

wherein, TP represents the true positive example, FN represents the false negative example, FP represents the false positive example to represent different prediction results, the accuracy reflects the accuracy of the model, the recall rate reflects the comprehensiveness of the model, the similarity entropy reflects the similarity of two sets, X represents TP + FP, Y represents TP + FN, the accuracy and the recall rate are comprehensively considered, the higher the three results are, and the better the effect is.