Semi-supervised MRI image tissue segmentation method and imaging method
By building a semi-supervised MRI image segmentation model containing a shared encoder and convolutional attention module, combined with image augmentation and loss function optimization, the problem of insufficient pseudo-label and consistency regularization in the existing methods is solved, and high accuracy and reliability of MRI image organization segmentation is achieved.
Patent Information
- Application Number
- CN202310224096.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-10
- Publication Date
- 2025-07-25
- Estimated Expiration
- 2043-03-10
AI Technical Summary
The existing semi-supervised medical image segmentation method has problems in MRI image tissue segmentation that pay too much attention to pseudo-label quality and ignore semantic information correlation, and the consistency regularization is too simple to lead to incorrect network learning and lead to degradation of segmentation performance.
Using a semi-supervised MRI image organization segmentation method, a model including a shared encoder, two decoders and a convolutional attention module is constructed, combined with image augmentation technology, supervised partial loss, boundary loss and unsupervised loss are calculated, image-level pseudo-labels are generated, and boundary optimization is performed to improve segmentation quality.
Accurate and robust organizational segmentation results on MRI images are achieved, which improves the reliability and accuracy of segmentation, and enhances the generalization ability and boundary segmentation quality of the network.
Smart Images
Figure CN116524178B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of image processing, and particularly relates to a semi - supervised MRI image tissue segmentation method and an imaging method. Background Art
[0002] With the development of economic technology and the improvement of people's living standards, people pay more and more attention to medical care and health. With the advent of the information age and the intelligent age, the integration of deep learning technology and computer - aided medical image processing technology has become the research focus of researchers.
[0003] MRI images are an important part of the medical field; therefore, the accuracy of tissue segmentation of MRI images will greatly affect the subsequent use and analysis of image data. At present, there are still two problems in tissue segmentation based on MRI images. First, although MRI supports multi - planar imaging and has high resolution, due to different parameter settings of MRI scanning equipment and morphological differences of tissues, it is still difficult to construct a general MRI tissue segmentation framework. Second, it is difficult to obtain labels in medical image segmentation, and high - quality labels often require the efforts of a large number of experienced doctors and experts. Therefore, semi - supervised medical image segmentation methods have emerged as the times require.
[0004] There are currently some semi - supervised medical image segmentation methods, which can be roughly divided into methods based on pseudo - label refinement and methods based on consistency regularization. Among them, for methods based on pseudo - label refinement, these methods focus too much on the quality of pseudo - labels and ignore the semantic information correlation between different images or views. For methods based on consistency regularization, these methods only focus on the uncertainty between different scales, and the fusion mechanism is too simple. In addition, some researchers have proposed semi - supervised medical image segmentation methods that jointly use pseudo - label refinement and consistency regularization. However, these methods often impose consistency constraints too directly and do not filter out uncertain pixels, which may cause the network to learn incorrect pixels and ignore correct pixels, making the learning of the network lose its goal and the performance decline accordingly. Summary of the Invention
[0005] One of the purposes of the present invention is to provide a semi - supervised MRI image tissue segmentation method with high reliability, good accuracy and good segmentation effect.
[0006] Another purpose of the present invention is to provide an imaging method including the semi - supervised MRI image tissue segmentation method described above.
[0007] The semi - supervised MRI image tissue segmentation method provided by the present invention includes the following steps:
[0008] S1. Obtain existing MRI image data and perform labeling to obtain labeled data images and unlabeled data images;
[0009] S2. Construct a preliminary model for semi - supervised MRI image tissue segmentation;
[0010] S3. Randomly select several images from the data obtained in step S1;
[0011] S4. Perform image augmentation operations on the images obtained in step S3 to obtain strongly augmented images and weakly augmented images;
[0012] S5. Among the images obtained in step S4, select the labeled images and input them into the current segmentation model, and calculate the supervised part loss and the boundary loss of the labeled images;
[0013] S6. Among the images obtained in step S4, select the unlabeled images and input them into the current segmentation model, and calculate the unsupervised loss;
[0014] S7. Combine the supervised part loss and boundary loss obtained in step S5 and the unsupervised loss obtained in step S6 to form a total loss function, and perform backpropagation through the gradient descent algorithm to update the parameters of the current segmentation model;
[0015] S8. Repeat steps S3 - S7 until the set conditions are met to obtain the final semi - supervised MRI image tissue segmentation model;
[0016] S9. Use the semi - supervised MRI image tissue segmentation model obtained in step S8 to perform actual tissue segmentation of MRI images.
[0017] The step of obtaining existing MRI image data and performing labeling in step S1 to obtain labeled data images and unlabeled data images specifically includes the following steps:
[0018] Obtain existing MRI image data;
[0019] Perform labeling on the data to obtain labeled data images and unlabeled data images Among them, is the i - th labeled data image, y i is the label of the i - th labeled data image, N is the total number of labeled data images, is the i - th unlabeled data image, and M is the total number of unlabeled data images.
[0020] The step of constructing a preliminary model for semi - supervised MRI image tissue segmentation in step S2 specifically includes the following steps:
[0021] The model includes a shared encoder, two decoders, and two convolutional attention modules;
[0022] The shared encoder includes five convolutional modules, each of which includes a 3×3 convolutional block, a batch normalization layer, and an activation function ReLU, and each convolutional module is followed by a 2×2 pooling layer;
[0023] The first decoder includes five upsampling modules, each of which includes an upsampling layer based on bilinear interpolation, a 3×3 convolutional block, a batch normalization layer, and an activation function ReLU;
[0024] The second decoder includes five upsampling modules, each of which includes an upsampling layer based on bilinear interpolation, a 3×3 convolutional block, a batch normalization layer, and an activation function ReLU;
[0025] The output of each layer of the shared encoder is concatenated with the same-sized features of the corresponding layer output of the first decoder to obtain the first concatenated feature, and the first concatenated feature is input into the first convolutional attention module;
[0026] The output of each layer of the shared encoder is concatenated with the same-sized features of the corresponding layer output of the second decoder to obtain the second concatenated feature, and the second concatenated feature is input into the second convolutional attention module;
[0027] The described convolutional attention module includes two parallel channel attention sub-modules and a spatial attention sub-module; the channel attention sub-module includes a max pooling layer, an average pooling layer, and a multi-layer perceptron; the spatial attention sub-module includes a max pooling layer, an average pooling layer, and a convolutional module layer, and the convolutional module layer includes a 3×3 convolutional block, a 2×2 pooling layer, a batch normalization layer, and an activation function ReLU;
[0028] The output features of the channel attention module, the output features of the spatial attention module are multiplied by the concatenated features and then added together, and input into a convolutional module to obtain the features activated by attention; finally, the features activated by attention are input into the corresponding network layer of the first decoder or the second decoder for subsequent decoding; the convolutional module includes a 3×3 convolutional block, a 2×2 pooling layer, a batch normalization layer, and an activation function ReLU.
[0029] The image augmentation operation is performed on the image obtained in step S3 in step S4 to obtain a strongly augmented image and a weakly augmented image, which specifically includes the following steps:
[0030] The image obtained in step S3 is divided into two parts:
[0031] The first part of the image is subjected to weak augmentation operations to obtain weakly augmented images; the weak augmentation operations include randomly resizing, cropping, flipping, and rotating;
[0032] The second part of the image is subjected to strong augmentation operations to obtain strongly augmented images; the strong augmentation operations include adjusting the image brightness, adjusting the image contrast, adjusting the image blurriness, and adding Gaussian noise;
[0033] Both the first part of the image and the second part of the image include labeled data images and unlabeled data images.
[0034] In the image obtained in step S4 as described in step S5, select the labeled images and input them into the current segmentation model, and calculate the supervised part loss and the boundary loss of the labeled images, which specifically include the following steps:
[0035] Input the labeled images into the current segmentation model to obtain the tissue segmentation results of the MRI images, and calculate the supervised part loss;
[0036] For the obtained tissue segmentation results of the MRI images, use the Sobel operator to extract the boundaries of the MRI images, then calculate the distance map of the images, and calculate the boundary loss.
[0037] The calculation of the supervised part loss specifically includes the following steps:
[0038] The supervised part loss L is calculated using the following formula sup :
[0039]
[0040] where D l is the set of labeled images; |D l | is the number of labeled images; x l is a labeled image; y is the label corresponding to x l ; l ce () is the cross-entropy loss function; A1() is the processing function for weakly augmented image processing; A2() is the processing function for strongly augmented image processing; f() is the processing function of the encoder; f(A1(x l )) is the feature map obtained after the weakly augmented image is processed by the encoder; f(A2(x l )) is the feature map obtained after the strongly augmented image is processed by the encoder; h1() is the processing function of the first decoder; h2() is the processing function of the second decoder.
[0041] The calculation of the distance map of the images specifically includes the following steps:
[0042] The distance map Dist(x, y) is calculated using the following formula:
[0043]
[0044] In the formula, B1() is the boundary map of the label; (x0, y0) is the position of the boundary pixel in the label; Euclid(a, b) is the Euclidean distance calculation function between position a and position b.
[0045] The calculation of the boundary loss specifically includes the following steps:
[0046] The boundary loss L is calculated using the following formula boundary :
[0047]
[0048] In the formula, C is the number of categories; p(x, y) is the probability map; B2 is the boundary map of the extracted MRI image.
[0049] In the image obtained in step S4 described in step S6, select the unlabeled image and input it into the current segmentation model, and calculate the unsupervised loss, which specifically includes the following steps:
[0050] Select the unlabeled image and input it into the current segmentation model to obtain the first view-level MRI tissue segmentation probability map p1 output by the first decoder and the second view-level MRI tissue segmentation probability map p2 output by the second decoder;
[0051] According to the obtained probability maps p1 and p2, calculate the corresponding confidence maps c1 and c2;
[0052] Generate image-level pseudo-labels based on the obtained probability maps p1 and p2 and confidence maps c1 and c2;
[0053] Calculate the unsupervised loss based on the obtained image-level pseudo-labels and probability maps.
[0054] The generation of the image-level pseudo-labels specifically includes the following steps:
[0055] Calculate the image-level pseudo-labels as
[0056] The calculation of the unsupervised loss specifically includes the following steps:
[0057] The unsupervised loss L is calculated using the following formula u :
[0058]
[0059] In the formula, Du is a collection of unlabeled images; |D u | is the number of unlabeled images; x u For unlabeled images.
[0060] Step S7 combines the supervised partial loss and boundary loss obtained in step S5 and the unsupervised loss obtained in step S6 to form a total loss function, which specifically includes the following steps:
[0061] The total loss function L constructed loss For L loss =L sup +L boundary +L u .
[0062] The present invention also discloses an imaging method including the semi-supervised MRI image tissue segmentation method, which specifically includes the following steps:
[0063] A. Obtain actual MRI image data;
[0064] B. using the semi-supervised MRI image tissue segmentation method to perform tissue segmentation on the MRI image data obtained in step A to obtain a tissue segmentation result;
[0065] C. The tissue segmentation result obtained in step B is marked and re-imaged on the MRI image obtained in step A, thereby completing the corresponding MRI image imaging.
[0066] The semi-supervised MRI image tissue segmentation method and imaging method provided by the present invention use the constructed MRI image tissue segmentation model to perform view-level prediction, then generate image-level pseudo labels to establish bidirectional consistency, and simultaneously perform boundary optimization to improve boundary segmentation quality. Therefore, the present invention can obtain accurate and robust tissue segmentation results on multi-task MRI images, and has high reliability, good precision and good segmentation effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0067] Figure 1 Schematic diagram of the process flow of the segmentation method of the present invention.
[0068] Figure 2 It is a schematic diagram of the overall flow of the segmentation method of the present invention.
[0069] Figure 3 Schematic diagram of the structure of the convolutional attention module in the segmentation method of the present invention.
[0070] Figure 4 It is a schematic diagram of the effect of the segmentation method of the present invention.
[0071] Figure 5Schematic diagram of the method flow of the imaging method of the present invention. Detailed implementation manners
[0072] As Figure 1 shown is the schematic diagram of the method flow of the segmentation method of the present invention, and its segmentation process is as Figure 2 shown:
[0073] The semi-supervised MRI image tissue segmentation method provided by the present invention includes the following steps:
[0074] S1. Obtain existing MRI image data and perform labeling to obtain labeled data images and unlabeled data images; specifically, it includes the following steps:
[0075] Obtain existing MRI image data;
[0076] Perform labeling on the data to obtain labeled data images and unlabeled data images Among them, is the i-th labeled data image, y i is the label of the i-th labeled data image, N is the total number of labeled data images, is the i-th unlabeled data image, and M is the total number of unlabeled data images;
[0077] Specifically in implementation, the obtained MRI images can be mixed images or single images: single images, that is, MRI images of only one tissue part, such as MRI data images of the female pelvis, MRI data images of the male prostate, or MRI data images of the heart, etc.; while mixed images can include images of multiple tissue parts at the same time, such as including MRI data images of the female pelvis, MRI data images of the male prostate, and MRI data images of the heart at the same time;
[0078] S2. Construct a preliminary model for semi-supervised MRI image tissue segmentation; specifically, it includes the following steps:
[0079] The model includes a shared encoder, two decoders, and two convolutional attention modules;
[0080] The shared encoder includes five convolutional modules, and each convolutional module includes a 3×3 convolutional block, a batch normalization layer, and an activation function ReLU, and each convolutional module is followed by a 2×2 pooling layer;
[0081] The first decoder includes five upsampling modules, and each upsampling module includes an upsampling layer based on bilinear interpolation, a 3×3 convolutional block, a batch normalization layer, and an activation function ReLU;
[0082] The second decoder includes five upsampling modules, and each upsampling module includes an upsampling layer based on bilinear interpolation, a 3×3 convolutional block, a batch normalization layer, and an activation function ReLU;
[0083] The output of each layer of the shared encoder is concatenated with the feature of the same size of the corresponding layer output of the first decoder to obtain the first concatenated feature, and the first concatenated feature is input into the first convolutional attention module; Specifically, in implementation, the output of the first convolutional module of the shared encoder and the output of the first upsampling module of the first decoder are concatenated, and the concatenated feature is input into the first convolutional attention module; Similarly, the output of the second convolutional module of the shared encoder and the output of the second upsampling module of the first decoder are concatenated, and the concatenated feature is input into the first convolutional attention module; And so on;
[0084] The output of each layer of the shared encoder is concatenated with the feature of the same size of the corresponding layer output of the second decoder to obtain the second concatenated feature, and the second concatenated feature is input into the second convolutional attention module; Specifically, in implementation, the output of the first convolutional module of the shared encoder and the output of the first upsampling module of the second decoder are concatenated, and the concatenated feature is input into the second convolutional attention module; Similarly, the output of the second convolutional module of the shared encoder and the output of the second upsampling module of the second decoder are concatenated, and the concatenated feature is input into the second convolutional attention module; And so on;
[0085] The convolutional attention module includes two parallel channel attention sub-modules and a spatial attention sub-module; The channel attention sub-module includes a max pooling layer, an average pooling layer, and a multi-layer perceptron; The spatial attention sub-module includes a max pooling layer, an average pooling layer, and a convolutional module layer, and the convolutional module layer includes a 3×3 convolutional block, a 2×2 pooling layer, a batch normalization layer, and an activation function ReLU;
[0086] The output feature of the channel attention module, the output feature of the spatial attention module are multiplied by the concatenated feature and then added together, and input into a convolutional module to obtain the feature activated by attention; The convolutional module includes a 3×3 convolutional block, a 2×2 pooling layer, a batch normalization layer, and an activation function ReLU; Finally, the feature activated by attention (i.e., the output feature of the convolutional attention module) is input into the corresponding network layer of the first decoder or the second decoder for subsequent decoding;
[0087] In specific implementation, taking the first convolutional attention module as an example: If the input feature is the concatenated feature of the output of the first convolutional module of the encoder and the output of the first upsampling module of the first decoder, then the output feature of the first convolutional attention module will be input into the second upsampling module of the first decoder as the input of the second upsampling module of the first decoder; If the input feature is the concatenated feature of the output of the second convolutional module of the encoder and the output of the second upsampling module of the first decoder, then the output feature of the first convolutional attention module will be input into the third upsampling module of the first decoder as the input of the third upsampling module of the first decoder; and so on; Meanwhile, the processing flow of the second convolutional attention module is the same as that of the first convolutional attention module;
[0088] The encoder can extract high-level features of the original data and reduce the dimension of the data;
[0089] The first and second decoders can gradually repair the details and spatial dimensions of the data to obtain the target output result;
[0090] The first and second convolutional attention modules can enable the network to focus on training key features, ignore non-key features, and improve the network segmentation accuracy;
[0091] The channel attention sub-module is used to generate a channel attention map, which can highlight key channels and suppress non-key channels, enabling the network to effectively learn global features and improve the network representation ability;
[0092] The spatial attention sub-module is used to generate a spatial attention map, which complements the channel attention sub-module, can highlight key spatial positions and suppress non-key spatial positions, enabling the network to effectively learn local features and improve the network performance;
[0093] Among them, the structure of the convolutional attention module adopted is as Figure 3 shown;
[0094] S3. Randomly select several images from the data obtained in step S1;
[0095] S4. Perform image augmentation operations on the images obtained in step S3 to obtain strongly augmented images and weakly augmented images; specifically including the following steps:
[0096] Divide the images obtained in step S3 into two parts:
[0097] Perform weak augmentation operations on the first part of the images to obtain weakly augmented images; The weak augmentation operations include randomly adjusting the size, cropping, flipping, and rotating;
[0098] Perform strong augmentation operations on the second part of the images to obtain strongly augmented images; the strong augmentation operations include adjusting the image brightness, adjusting the image contrast, adjusting the image blurriness, and adding Gaussian noise;
[0099] Both the first part of the images and the second part of the images include labeled data images and unlabeled data images;
[0100] S5. Among the images obtained in step S4, select the labeled images and input them into the current segmentation model, and calculate the supervised part loss and the boundary loss of the labeled images; specifically, it includes the following steps:
[0101] Input the labeled images into the current segmentation model to obtain the tissue segmentation results of the MRI images, and calculate the supervised part loss;
[0102] Specifically in implementation, calculating the supervised part loss specifically includes the following steps:
[0103] Calculate the supervised part loss L using the following formula sup :
[0104]
[0105] In the formula, D l is the set of labeled images; |D l | is the number of labeled images; x l is a labeled image; y is the label corresponding to x l ; l ce () is the cross-entropy loss function; A1() is the processing function for weakly augmented image processing; A2() is the processing function for strongly augmented image processing; f() is the processing function of the encoder; f(A1(x l )) is the feature map obtained after the weakly augmented image is processed by the encoder; f(A2(x l )) is the feature map obtained after the strongly augmented image is processed by the encoder; h1() is the processing function of the first decoder; h2() is the processing function of the second decoder;
[0106] For the obtained tissue segmentation results of the MRI images, use the Sobel operator to extract the boundaries of the MRI images, then calculate the distance map of the images, and calculate the boundary loss;
[0107] Specifically in implementation, calculating the distance map of the images specifically includes the following steps:
[0108] Calculate the distance map Dist(x, y) using the following formula:
[0109]
[0110] where B1() is the boundary map of the label; (x0, y0) is the position of the boundary pixel in the label; Euclid(a, b) is the Euclidean distance calculation function between position a and position b;
[0111] The calculation of the boundary loss specifically includes the following steps:
[0112] The boundary loss L is calculated using the following formula boundary :
[0113]
[0114] where C is the number of categories; p(x, y) is the probability map; B2 is the boundary map of the extracted MRI image;
[0115] S6. In the image obtained in step S4, select the unlabeled images and input them into the current segmentation model, and calculate the unsupervised loss; specifically, it includes the following steps:
[0116] Select the unlabeled images and input them into the current segmentation model to obtain the first view-level MRI tissue segmentation probability map p1 output by the first decoder and the second view-level MRI tissue segmentation probability map p2 output by the second decoder;
[0117] According to the obtained probability maps p1 and p2, calculate the corresponding confidence maps c1 and c2;
[0118] According to the obtained probability maps p1 and p2 and confidence maps c1 and c2, generate image-level pseudo-labels; specifically, it includes the following steps:
[0119] Calculate the image-level pseudo-labels as
[0120] According to the obtained image-level pseudo-labels and probability maps, calculate the unsupervised loss; specifically, it includes the following steps:
[0121] The unsupervised loss L is calculated using the following formula u :
[0122]
[0123] where D u is the set of unlabeled images; |D u | is the number of unlabeled images; x u is an unlabeled image;
[0124] S7. Combine the supervised partial loss and boundary loss obtained in step S5 and the unsupervised loss obtained in step S6 to form the total loss function, and perform backpropagation through the gradient descent algorithm to update the parameters of the current segmentation model;
[0125] In specific implementation, the composition of the total loss function specifically includes the following steps:
[0126] The constructed total loss function L loss is L loss = L sup + L boundary + L u ;
[0127] S8. Repeat steps S3 - S7 until the set conditions are met to obtain the final semi - supervised MRI image tissue segmentation model;
[0128] S9. Use the semi - supervised MRI image tissue segmentation model obtained in step S8 to perform actual tissue segmentation of MRI images.
[0129] The present invention proposes an image tissue segmentation method, which inputs dual views into a single encoder and dual decoders to obtain view - level predictions, and then uses the confidence map and probability map to fuse them to generate better image - level pseudo - labels, thereby generating robust and accurate segmentation results;
[0130] The present invention improves the convolutional attention module, parallelly uses the channel attention sub - module and the spatial attention sub - module, enables the network to learn channel features and spatial features simultaneously, and enhances the generalization ability of the network;
[0131] Moreover, the present invention proposes a boundary optimization module, which uses the Sobel operator to extract the boundary of the MRI image, and then calculates the distance map of the image as the boundary soft label, improving the quality of boundary segmentation.
[0132] The following presents an embodiment to further illustrate the segmentation effect of the present invention:
[0133] The segmentation method proposed by the present invention is compared with existing segmentation methods. When comparing, a female pelvic MRI dataset with 134 training sets and 37 test sets, a prostate MRI dataset with 497 training sets and 125 test sets, and a cardiac MRI dataset with 690 training sets and 173 test sets are used. The fully supervised method uses all labels of the training set for training, and other methods all use 10% of the labeled data. The average intersection - over - union ratio, Dice coefficient, and Hausdorff distance between the segmentation result and the Ground Truth are used as evaluation criteria.
[0134] The specific results are shown in Tables 1 - 3:
[0135] Table 1 Data table of female pelvic MRI tissue segmentation results
[0136] Segmentation method Average Intersection over Union (%) Dice coefficient (%) Hausdorff distance (mm) Fully supervised method 74.18 81.77 4.28 Semi-supervised method in 2017 76.45 83.81 3.72 Semi-supervised method in 2021 77.17 84.50 3.81 Method of the present invention 78.28 85.21 3.53
[0137] Table 2 Data Table of Prostate MRI Tissue Segmentation Results
[0138] Segmentation method Average Intersection over Union (%) Dice coefficient (%) Hausdorff distance (mm) Fully supervised method 84.58 89.13 4.71 Semi-supervised method in 2017 85.14 89.61 4.45 Semi-supervised method in 2021 87.07 91.50 4.55 Method of the present invention 87.78 92.17 4.26
[0139] Table 3 Data Table of Heart MRI Tissue Segmentation Results
[0140] Segmentation method Average Intersection over Union (%) Dice coefficient (%) Hausdorff distance (mm) Fully supervised method 63.14 69.07 3.63 Semi-supervised method in 2017 72.59 78.53 3.05 Semi-supervised method in 2021 73.63 79.70 3.00 Method of the present invention 74.46 79.84 2.87
[0141] As can be seen from Tables 1 to 3, the segmentation method proposed in the present invention is superior to the fully supervised method, the classical semi-supervised method, and the latest method in all three indicators, and the segmentation method of the present invention can obtain more accurate segmentation results.
[0142] Such as Figure 4 shown is the effect schematic diagram of the segmentation method of the present invention: Figure 4 The results of the semi-supervised method and the segmentation method of the present invention are compared. From left to right, they are the original image, GroundTruth, the semi-supervised method in 2017, the semi-supervised method in 2021, and the segmentation result of the segmentation method of the present invention; from Figure 4 it can be seen that the segmentation method of the present invention performs better in the integrity of MRI tissue segmentation and the continuity of the narrow part.
[0143] Such as Figure 5 shown is the method flow schematic diagram of the imaging method of the present invention: The imaging method including the semi-supervised MRI image tissue segmentation method disclosed in the present invention specifically includes the following steps:
[0144] A. Obtain actual MRI image data;
[0145] B. Adopt the semi-supervised MRI image tissue segmentation method described above to perform tissue segmentation on the MRI image data obtained in step A to obtain tissue segmentation results;
[0146] C. Mark and perform secondary imaging on the tissue segmentation results obtained in step B on the MRI image obtained in step A, thereby completing the corresponding MRI image imaging.
[0147] In specific implementation, the imaging method of the present invention can be used in existing MRI image acquisition devices, such as MRI machines. When specifically used, the imaging method of the present invention is integrated into the existing MRI image acquisition device, and then the original MRI image is acquired using the existing technology. Then, the acquired original MRI image is subjected to secondary imaging using the described imaging method, and an MRI image with tissue segmentation results is obtained and directly output. In this way, medical workers (including clinicians, radiologists, or experimental personnel, etc.) can obtain MRI images directly with tissue segmentation results, which will greatly facilitate the existing users.
Claims
1. A semi - supervised MRI image tissue segmentation method, comprising the following steps: S1. Obtain existing MRI image data and perform labeling to obtain labeled data images and unlabeled data images; S2. Construct a preliminary semi - supervised MRI image tissue segmentation model; S3. Randomly select several images from the data obtained in step S1; S4. Perform image augmentation operations on the images obtained in step S3 to obtain strongly augmented images and weakly augmented images; S5. Among the images obtained in step S4, select the labeled images and input them into the current segmentation model, and calculate the supervised part loss and the boundary loss of the labeled images; S6. Among the images obtained in step S4, select the unlabeled images and input them into the current segmentation model, and calculate the unsupervised loss; S7. Combine the supervised part loss and boundary loss obtained in step S5 and the unsupervised loss obtained in step S6 to form a total loss function, and perform backpropagation through the gradient descent algorithm to update the parameters of the current segmentation model; S8. Repeat steps S3 - S7 until the set conditions are met to obtain the final semi - supervised MRI image tissue segmentation model; S9. Use the semi - supervised MRI image tissue segmentation model obtained in step S8 to perform actual tissue segmentation of MRI images.
2. The semi-supervised MRI image tissue segmentation method according to claim 1, characterized in that The step of obtaining existing MRI image data and performing labeling in step S1 to obtain labeled data images and unlabeled data images specifically includes the following steps: Obtain existing MRI image data; Label the data to obtain the labeled data images and the unlabeled data images where is the i-th labeled data image, and y i is the label of the i-th labeled data image, N is the total number of labeled data images, is the i-th unlabeled data image, and M is the total number of unlabeled data images.
3. The semi-supervised MRI image tissue segmentation method according to claim 2, characterized in that The step of constructing a preliminary semi - supervised MRI image tissue segmentation model in step S2 specifically includes the following steps: The model includes a shared encoder, two decoders, and two convolutional attention modules; The shared encoder includes five convolutional modules, each convolutional module includes a 3×3 convolutional block, a batch normalization layer, and an activation function ReLU, and each convolutional module is followed by a 2×2 pooling layer; The first decoder includes five upsampling modules, each upsampling module includes an upsampling layer based on bilinear interpolation, a 3×3 convolutional block, a batch normalization layer, and an activation function ReLU; The second decoder includes five upsampling modules, each upsampling module includes an upsampling layer based on bilinear interpolation, a 3×3 convolutional block, a batch normalization layer, and an activation function ReLU; The output of each layer of the shared encoder is concatenated with the same - sized features of the corresponding layer output of the first decoder to obtain the first concatenated feature, and the first concatenated feature is input into the first convolutional attention module; The output of each layer of the shared encoder is concatenated with the same - sized features of the corresponding layer output of the second decoder to obtain the second concatenated feature, and the second concatenated feature is input into the second convolutional attention module; The convolutional attention module includes two parallel channel attention sub - modules and spatial attention sub - modules; Among them, the channel attention sub-module includes a max pooling layer, an average pooling layer, and a multi-layer perceptron; the spatial attention sub-module includes a max pooling layer, an average pooling layer, and a convolutional module layer, and the convolutional module layer includes a 3×3 convolutional block, a 2×2 pooling layer, a batch normalization layer, and an activation function ReLU; The output features of the channel attention module, the output features of the spatial attention module are multiplied and then added to the concatenated features, and then input into a convolutional module to obtain the features activated by attention; finally, the features activated by attention are input into the corresponding network layer of the first decoder or the second decoder for subsequent decoding; the convolutional module includes a 3×3 convolutional block, a 2×2 pooling layer, a batch normalization layer, and an activation function ReLU.
4. The semi-supervised MRI image tissue segmentation method according to claim 3, wherein The image augmentation operation is performed on the image obtained in step S3 in step S4 to obtain a strongly augmented image and a weakly augmented image, which specifically includes the following steps: The image obtained in step S3 is divided into two parts: The first part of the image is subjected to a weak augmentation operation to obtain a weakly augmented image; the weak augmentation operation includes randomly adjusting the size, cropping, flipping, and rotating; The second part of the image is subjected to a strong augmentation operation to obtain a strongly augmented image; the strong augmentation operation includes adjusting the image brightness, adjusting the image contrast, adjusting the image blurriness, and adding Gaussian noise; Both the first part of the image and the second part of the image include labeled data images and unlabeled data images.
5. The semi-supervised MRI image tissue segmentation method according to claim 4, wherein In the image obtained in step S4 in step S5, the labeled image is selected and input into the current segmentation model, and the supervised part loss and the boundary loss of the labeled image are calculated, which specifically includes the following steps: The labeled image is input into the current segmentation model to obtain the tissue segmentation result of the MRI image, and the supervised part loss is calculated; For the obtained tissue segmentation result of the MRI image, the Sobel operator is used to extract the boundary of the MRI image, then the distance map of the image is calculated, and the boundary loss is calculated.
6. The semi-supervised based MRI image tissue segmentation method according to claim 5, wherein The calculation of the supervised part loss specifically includes the following steps: The supervised part loss L is calculated using the following formula sup : where D l is a set of labeled images; D l is the number of labeled images; x l is a labeled image; y is the label corresponding to x l ; l ce () is the cross-entropy loss function; A1() is the processing function for weakly augmented image processing; A2() is the processing function for strongly augmented image processing; f() is the processing function of the encoder; f(A1(x l )) is the feature map obtained by processing the weakly augmented image through the encoder; f(A2(x l )) is the feature map obtained by processing the strongly augmented image through the encoder; h1() is the processing function of the first decoder; h2() is the processing function of the second decoder; The calculation of the distance map of the image specifically includes the following steps: The distance map Dist(x,y) is calculated using the following formula: In the formula, B1() is the boundary map of the label; (x0,y0) is the position of the boundary pixel in the label; Euclid(a,b) is the Euclidean distance calculation function between position a and position b; The calculation of the obtained boundary loss specifically includes the following steps: The boundary loss L is calculated using the following formula boundary :[[]]END]] In the formula, C is the number of categories; p(x,y) is the probability map; B2 is the boundary map of the extracted MRI image.
7. The semi-supervised MRI image tissue segmentation method according to claim 6, characterized in that In the image obtained in step S4 in step S6, the unlabeled image is selected and input into the current segmentation model, and the unsupervised loss is calculated, which specifically includes the following steps: Select the unlabeled image and input it into the current segmentation model to obtain the first view-level MRI tissue segmentation probability map p1 output by the first decoder and the second view-level MRI tissue segmentation probability map p2 output by the second decoder; According to the obtained probability maps p1 and p2, the corresponding confidence maps c1 and c2 are calculated; Generate image-level pseudo-labels based on the obtained probability maps p1 and p2 and confidence maps c1 and c2. Calculate the unsupervised loss based on the obtained image-level pseudo-labels and probability maps.
8. The semi-supervised MRI image tissue segmentation method according to claim 7, characterized in that The generation of the image-level pseudo-labels specifically includes the following steps: Calculate the image-level pseudo-labels For The calculation of the unsupervised loss specifically includes the following steps: The unsupervised loss L is calculated using the following formula u :[[]]END]] where D u is a set of unlabeled images; D u is the number of unlabeled images; x u is an unlabeled image.
9. The semi-supervised MRI image tissue segmentation method according to claim 8, characterized in that The comprehensive supervised partial loss and boundary loss obtained in step S5 and the unsupervised loss obtained in step S6 in step S7 constitute the total loss function, specifically including the following steps: The total loss function L constructed loss is L loss = L sup + L boundary + L u .
10. An imaging method including the semi-supervised based MRI image tissue segmentation method according to any one of claims 1 to 9, specifically including the following steps: A. Obtain actual MRI image data. B. Use the semi-supervised based MRI image tissue segmentation method according to any one of claims 1 to 9 to perform tissue segmentation on the MRI image data obtained in step A to obtain a tissue segmentation result. C. Mark and re-image the tissue segmentation result obtained in step B on the MRI image obtained in step A, thereby completing the corresponding MRI image imaging.