Heart nuclear magnetic resonance image segmentation method of expanded convolutional network based on direction field features
Through the expanded convolutional network based on directional field features, the problems of feature information loss and invalid information redundancy in cardiac MRI image segmentation are solved, automatic segmentation is realized, segmentation accuracy and efficiency are improved, and it is suitable for clinical needs in underdeveloped areas. .
Patent Information
- Application Number
- CN202310635490.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-01
- Publication Date
- 2024-12-03
AI Technical Summary
The existing cardiac MRI image segmentation methods have problems such as loss of feature information, redundancy of invalid information, blurred MRI artifacts and inhomogeneity, resulting in blurred inter-classifications and unclear boundaries, and lack of automated solutions, which affects the accuracy of doctors' decision-making. Especially in underdeveloped areas, it is even more serious.
Using an expanded convolution network based on directional field features, multi-scale hollow convolution module and directional field module are used to extract and utilize multi-scale feature information to reduce the loss of feature information and the increase of useless information, and generate more accurate segmentation results.
It realizes automated segmentation of different parts of cardiac MRI images, replaces the doctor's manual segmentation, improves the accuracy and efficiency of segmentation, reduces the loss of feature information, enhances inter-class differences and intra-class similarity, and generates more accurate Final segmentation result.
Smart Images

Figure CN119067909A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of translational medicine and artificial intelligence technology, and in particular to an automatic segmentation method for cardiac magnetic resonance images. Background Art
[0002] Cardiovascular disease has attracted widespread attention in the medical community due to its hazards. With the development of cardiac imaging technology, medical personnel have been able to further study the disease. In the diagnosis of this disease, short-axis cardiac magnetic resonance imaging (MRI) is widely used due to its non-invasive imaging characteristics. In clinical cardiology, clinicians need to distinguish the left ventricle (LV), right ventricle (RV) and myocardium (MYO) from short-axis cardiac MRI. Manual identification of different parts of the heart is a time-consuming, tedious and vulnerable task to external influences. Methods for automatically performing cardiac MRI segmentation tasks have become necessary, which allows inexperienced people to easily perform segmentation work. At the same time, segmentation methods can reduce repetition and decompose long sentences to improve text clarity, conciseness and readability. Although traditional MRI segmentation methods have made some progress, there are still many problems, such as:
[0003] (1) Existing networks usually use ordinary convolutional networks. However, using such networks can easily lead to the loss of feature information or the addition of too much invalid information, making it impossible to fully utilize the feature information;
[0004] (2) Some methods do not consider the blurriness and inhomogeneity of MRI artifacts, which may lead to problems such as blurring between categories and unclear boundaries;
[0005] (3) Some convolutional networks do not properly utilize feature information of different scales;
[0006] There is still a lack of doctors with rich experience in MRI segmentation, which is more serious in underdeveloped areas. On the other hand, the accuracy of cardiac MRI segmentation will directly affect doctors' decision-making. Summary of the invention
[0007] The present invention aims to overcome the shortcomings and deficiencies of the prior art in cardiac magnetic resonance image segmentation. The invention proposes a cardiac magnetic resonance image segmentation method based on a dilated convolutional network with directional field features, aiming to solve the problems of effective feature information being ignored and invalid feature redundancy in the existing methods.
[0008] The object of the present invention is achieved by the following technical solution: A cardiac magnetic resonance image segmentation method based on a dilated convolutional network with directional field features, comprising the following steps:
[0009] (1) acquiring multiple cardiac MRI images;
[0010] (2) inputting each cardiac MRI image obtained in step (1) into the trained neural network and outputting the result;
[0011] Preferably, the convolutional network structure is as follows:
[0012] The first layer is the input layer, which takes a cardiac MRI image of size 256*256*1, and converts its size to 256*256*64 and then passes it through a 2*2 maximum pooling layer. The final input size is 128*128*64.
[0013] The second layer uses a dilated convolution module to receive the input layer image. This layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively. The number of convolution kernels is 128, the step size is 1, the SAME method is used for padding, and the linear rectification function is used as the activation function. Finally, this layer passes through a 2*2 maximum pooling layer and outputs a matrix of size 64*64*128.
[0014] The third layer uses a dilated convolution module to receive the input layer image. This layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively, the number of convolution kernels is 256, the step size is 1, the SAME method is used for padding, and the linear rectification function is used as the activation function. Finally, this layer passes through a 2*2 maximum pooling layer and outputs a matrix of size 32*32*256;
[0015] The fourth layer uses a dilated convolution module to receive the input layer image. This layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively, the number of convolution kernels is 512, the step size is 1, the SAME method is used for padding, and the linear rectification function is used as the activation function. Finally, this layer passes through a 2*2 maximum pooling layer and outputs a matrix of size 16*16*512;
[0016] The fifth layer uses a dilated convolution module to receive the input layer image. This layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively, the number of convolution kernels is 512, the step size is 1, the SAME method is used for padding, and a linear rectification function is used as the activation function. Finally, this layer uses a bilinear interpolation function to enlarge the image by two times and outputs a matrix of size 32*32*512; The sixth layer uses a dilated convolution module to receive the input layer image and add it to the input of the fourth layer at the channel level, and outputs a matrix of size 32*32*1024. Then, this layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively, the number of convolution kernels is 256, the step size is 1, the SAME method is used for padding, and a linear rectification function is used as the activation function. Finally, this layer uses a bilinear interpolation function to enlarge the image by two times and outputs a matrix of size 64*64*256.
[0017] The seventh layer uses a dilated convolution module, which receives the input layer image and adds it to the input of the third layer at the channel level, outputting a matrix of size 64*64*512. Then, this layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively, the number of convolution kernels is 128, the step size is 1, the SAME method is used for padding, and a linear rectification function is used as the activation function. Finally, this layer uses a bilinear interpolation function to enlarge the image by two times and output a matrix of size 128*128*128.
[0018] The eighth layer uses a dilated convolution module, which receives the input layer image and adds it to the input of the second layer at the channel level, outputting a matrix of size 128*128*256. Then, this layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively, the number of convolution kernels is 64, the step size is 1, the SAME method is used for padding, and a linear rectification function is used as the activation function. Finally, this layer uses a bilinear interpolation function to enlarge the image by two times and output a matrix of size 256*256*64.
[0019] The ninth layer uses a dilated convolution module, which receives the input layer image and adds it to the input of the first layer at the channel level, outputting a matrix of size 256*256*128. Then, this layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively, the number of convolution kernels is 64, the step size is 1, the SAME method is used for padding, and a linear rectification function is used as the activation function. Finally, this layer uses a bilinear interpolation function to enlarge the image by two times and output a matrix of size 256*256*64.
[0020] The tenth layer is a feature aggregation module, which aggregates the features of the sixth, seventh, eighth, and ninth layers into a 256*256*512 matrix in the channel dimension. Then, this layer uses a 1*1 convolution kernel to reduce the number of channels. The number of convolution kernels is 64, the step size is 1, the SAME method is used for padding, and the linear rectification function is used as the activation function, and a matrix of 256*256*64 is output.
[0021] The eleventh layer is the direction field module, which receives the matrix of size 256*256*64 output from the tenth layer. This layer is composed of a 1*1 convolutional direction field module to learn the direction field. Its input is the final output feature of the model decoder, and it outputs a matrix of size 256*256*64. The background pixel of the direction field is (0, 0), which is defined as follows:
[0022]
[0023] Where a represents the foreground pixel, b represents the pixel closest to the cardiac tissue boundary, ba→ is the direction vector between b and a, the invention is normalized by distance, the direction field module provides a direction vector for each pixel to point to the central area, thereby predicting the relationship between pixels, after generating the direction field, the model uses the generated direction field F to improve the output characteristic M 0 To obtain the improved feature M N , the features in the center area are corrected to M 0 ,Each pixel is iteratively updated, and the operation is defined as follows:
[0024]
[0025] Where Ω is the image domain, k represents the kth step, N is the total number of iterations, and p x and p y are the x and y coordinates of pixel a respectively. Then, M N and M 0 Connect them together and use them as the input of the final classifier to generate the segmentation result;
[0026] The twelfth layer is the convolutional layer, which receives the feature matrix of size 256*256*64 output by the tenth layer and outputs a matrix of size 256*256*4. The four channels represent the three segmented parts and the background respectively.
[0027] Preferably, the convolutional network is trained by the following steps:
[0028] (3-1) Obtain a cardiac MRI image dataset and divide the dataset into a training set and a test set in a ratio of 8:2;
[0029] (3-2) Step (3-2) is to send the cardiac MRI image dataset obtained in step (3-1) to a physician for manual data annotation to generate manual data annotation results as the annotated cardiac MRI image dataset. In this step, the accuracy and consistency of the manual data annotation need to be ensured to improve the quality of the cardiac MRI image data.
[0030] (3-3) performing data enhancement on the training set in the annotated cardiac MRI image data set in step (3-2) to obtain an enhanced cardiac MRI image training set. During this process, the invention performs operations such as image rotation, flipping and scaling to obtain more abundant data samples, thereby improving the universality and performance of the model; (3-4) performing standardization and normalization on the cardiac MRI image training set after data enhancement in step (3-3) to obtain a preprocessed cardiac MRI image training set. Such processing steps can eliminate the deviation between data samples, reduce the differences between samples, and help improve the training effect and prediction accuracy of the model;
[0031] (3-5) inputting the cardiac MRI image in the preprocessed cardiac MRI image training set obtained in step (3-4) into the task multi-scale dilated convolutional neural network to obtain the corresponding reasoning output result, and inputting the reasoning output result and the cardiac MRI image annotated by the ultrasound physician in step (3-2) into the cross entropy function of the multi-scale dilated convolutional neural network to obtain the loss value;
[0032] (3-6) Using the adaptive moment estimation optimizer and the loss value obtained in step (3-5), the loss function of the convolutional network is optimized to iteratively update the parameters in the shallow convolutional network. This optimization strategy has the advantages of fast convergence and efficient processing of large amounts of data, which helps to accelerate model learning and improve the generalization performance of the model;
[0033] (3-7) Repeat the above steps (3-5) and (3-6) until the multi-scale dilated convolutional neural network converges to an optimal state, thereby obtaining a trained multi-scale dilated convolutional neural network;
[0034] Compared with the prior art, the present invention has the following beneficial effects:
[0035] (1) Since the present invention adopts steps (1) to (3), it realizes the automatic segmentation of different parts of cardiac MRI images through multi-scale hole convolutional neural network technology, which can replace doctors to complete the segmentation of cardiac MRI images. The present invention can solve the technical problems of traditional manual quality control with large workload and long time consumption;
[0036] (2) Since the present invention applies the dilated convolution technique in step (2), it can effectively extract and utilize multi-scale information, reducing the loss of feature information or the increase of useless information;
[0037] (3) Since the present invention applies the directional field technology in step (2) to enhance the differences between classes and the similarities within classes, it can improve the original network features using directional features and generate a more accurate final segmentation result. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 It is a flow chart of the cardiac MRI image segmentation method of the present invention.
[0039] Figure 2 This is an example of a result generated by the cardiac MRI image segmentation of the present invention, and the result image includes left ventricle, right ventricle and myocardial images (all on one image).
[0040] Figure 3 It is a multi-scale hole convolutional neural network model diagram of the cardiac MRI image segmentation method of the present invention.
[0041] Figure 4 It is a hollow convolution module in the cardiac MRI image segmentation method of the present invention. DETAILED DESCRIPTION
[0042] In order to more clearly describe the purpose, technical solutions and advantages of the present invention, this article provides a detailed description with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described here are only used to explain the present invention and are not used to limit the present invention.
[0043] like Figure 1 As shown, the present invention provides an automatic segmentation method for cardiac magnetic resonance images, comprising the following steps:
[0044] (1) Obtain cardiac MRI images;
[0045] Specifically, cardiac MRI image segmentation includes three parts: left ventricle, right ventricle, and myocardium. This method can simultaneously segment the three parts ( Figure 2 ).
[0046] (2) inputting each cardiac MRI image obtained in step (1) into the trained multi-scale dilated convolutional neural network to segment the cardiac MRI image into three different parts, and outputting the result;
[0047] Specifically, the multi-scale dilated convolutional neural network in this step contains 8 dilated convolution modules and 1 direction field module ( Figure 3 ), where the dilated convolution module contains three dilated convolutions, whose dilation rates are set to 1, 2, and 3 respectively ( Figure 4 ), the direction field module is a 1*1 convolution, and the structure of the multi-scale hole convolutional neural network is as follows:
[0048] The first layer is the input layer, which takes a cardiac MRI image of size 256*256*1, and converts its size to 256*256*64 and then passes it through a 2*2 maximum pooling layer. The final input size is 128*128*64.
[0049] The second layer uses a dilated convolution module to receive the input layer image. This layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively. The number of convolution kernels is 128, the step size is 1, the SAME method is used for padding, and the linear rectification function is used as the activation function. Finally, this layer passes through a 2*2 maximum pooling layer and outputs a matrix of size 64*64*128.
[0050] The third layer uses a dilated convolution module to receive the input layer image. This layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively, the number of convolution kernels is 256, the step size is 1, the SAME method is used for padding, and the linear rectification function is used as the activation function. Finally, this layer passes through a 2*2 maximum pooling layer and outputs a matrix of size 32*32*256;
[0051] The fourth layer uses a dilated convolution module to receive the input layer image. This layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively, the number of convolution kernels is 512, the step size is 1, the SAME method is used for padding, and the linear rectification function is used as the activation function. Finally, this layer passes through a 2*2 maximum pooling layer and outputs a matrix of size 16*16*512;
[0052] The fifth layer uses a dilated convolution module to receive the input layer image. This layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3, respectively. The number of convolution kernels is 512, the step size is 1, and the SAME method is used for padding and a linear rectification function is used as the activation function. Finally, this layer doubles the image through a bilinear interpolation function and outputs a matrix of size 32*32*512; The sixth layer uses a dilated convolution module to receive the input layer image and add it to the input of the fourth layer at the channel level, and outputs a matrix of size 32*32*1024. Then, this layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3, respectively. The number of convolution kernels is 256, the step size is 1, the SAME method is used for padding and a linear rectification function is used as the activation function. Finally, this layer doubles the image through a bilinear interpolation function and outputs a matrix of size 64*64*256.
[0053] The seventh layer uses a dilated convolution module, which receives the input layer image and adds it to the input of the third layer at the channel level, outputting a matrix of size 64*64*512. Then, this layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively, the number of convolution kernels is 128, the step size is 1, the SAME method is used for padding, and a linear rectification function is used as the activation function. Finally, this layer uses a bilinear interpolation function to enlarge the image by two times and output a matrix of size 128*128*128.
[0054] The eighth layer uses a dilated convolution module, which receives the input layer image and adds it to the input of the second layer at the channel level, outputting a matrix of size 128*128*256. Then, this layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively, the number of convolution kernels is 64, the step size is 1, the SAME method is used for padding, and a linear rectification function is used as the activation function. Finally, this layer uses a bilinear interpolation function to enlarge the image by two times and output a matrix of size 256*256*64.
[0055] The ninth layer uses a dilated convolution module, which receives the input layer image and adds it to the input of the first layer at the channel level, outputting a matrix of size 256*256*128. Then, this layer uses three consecutive dilated convolution kernels of size 3*3, with dilation rates of 1, 2, and 3 respectively, the number of convolution kernels is 64, the step size is 1, the SAME method is used for padding, and a linear rectification function is used as the activation function. Finally, this layer uses a bilinear interpolation function to enlarge the image by two times and output a matrix of size 256*256*64.
[0056] The tenth layer is a feature aggregation module, which aggregates the features of the sixth, seventh, eighth, and ninth layers into a 256*256*512 matrix in the channel dimension. Then, this layer uses a 1*1 convolution kernel to reduce the number of channels. The number of convolution kernels is 64, the step size is 1, the SAME method is used for padding, and the linear rectification function is used as the activation function, and a matrix of 256*256*64 is output.
[0057] The eleventh layer is the direction field module, which receives the matrix of size 256*256*64 output from the tenth layer. This layer is composed of a 1*1 convolutional direction field module to learn the direction field. Its input is the final output feature of the model decoder, and it outputs a matrix of size 256*256*64. The background pixel of the direction field is (0, 0), which is defined as follows:
[0058]
[0059] Where a represents the foreground pixel, b represents the pixel closest to the cardiac tissue boundary, ba→ is the direction vector between b and a, the invention is normalized by distance, the direction field module provides a direction vector for each pixel to point to the central area, thereby predicting the relationship between pixels, after generating the direction field, the model uses the generated direction field F to improve the output characteristic M 0 To obtain the improved feature M N , the features in the center area are corrected to M 0 ,Each pixel is iteratively updated, and the operation is defined as follows:
[0060]
[0061] Where Ω is the image domain, k represents the kth step, N is the total number of iterations, and p x and p y are the x and y coordinates of pixel a respectively. Then, M N and M 0 Connect them together and use them as the input of the final classifier to generate the segmentation result;
[0062] The twelfth layer is the convolutional layer, which receives the feature matrix of size 256*256*64 output by the tenth layer and outputs a matrix of size 256*256*4. The four channels represent the three segmented parts and the background respectively.
[0063] Preferably, the multi-scale dilated convolutional neural network is trained by the following steps:
[0064] (3-1) Obtain a cardiac MRI image dataset and divide the dataset into a training set and a test set in a ratio of 8:2;
[0065] (3-2) Step (3-2) is to send the cardiac MRI image dataset obtained in step (3-1) to a physician for manual data annotation to generate manual data annotation results as the annotated cardiac MRI image dataset. In this step, the accuracy and consistency of the manual data annotation need to be ensured to improve the quality of the cardiac MRI image data.
[0066] (3-3) performing data enhancement on the training set in the annotated cardiac MRI image data set in step (3-2) to obtain an enhanced cardiac MRI image training set. During this process, the invention performs operations such as image rotation, flipping and scaling to obtain more abundant data samples, thereby improving the universality and performance of the model; (3-4) performing standardization and normalization on the cardiac MRI image training set after data enhancement in step (3-3) to obtain a preprocessed cardiac MRI image training set. Such processing steps can eliminate the deviation between data samples, reduce the differences between samples, and help improve the training effect and prediction accuracy of the model;
[0067] (3-5) inputting the cardiac MRI image in the preprocessed cardiac MRI image training set obtained in step (3-4) into the task multi-scale dilated convolutional neural network to obtain the corresponding reasoning output result, and inputting the reasoning output result and the cardiac MRI image annotated by the ultrasound physician in step (3-2) into the cross entropy function of the multi-scale dilated convolutional neural network to obtain the loss value;
[0068] (3-6) Using the adaptive moment estimation optimizer and the loss value obtained in step (3-5), the loss function of the convolutional network is optimized to iteratively update the parameters in the shallow convolutional network. This optimization strategy has the advantages of fast convergence and efficient processing of large amounts of data, which helps to accelerate model learning and improve the generalization performance of the model.
[0069] (3-7) Repeat the above steps (3-5) and (3-6) until the multi-scale dilated convolutional neural network converges to an optimal state, thereby obtaining a trained multi-scale dilated convolutional neural network;
[0070] Performance Test Results
[0071] The following is a test set of representative cardiac MRI images randomly selected to test the proposed model. The test set includes 100 cardiac MRI images. The accuracy is used as the evaluation criterion to test the detection results of the multi-scale hole convolutional neural network. The accuracy of segmenting the left ventricle, right ventricle and myocardium is as high as 94.7%, 90.8% and 89.9% respectively, which can meet the current clinical needs for cardiac MRI images. Among them, the precise values of LV, RV and MYO have reached a relatively good standard. In the data set, for the three different parts of the heart, the LV segmentation accuracy is the highest, while the MYO segmentation accuracy is lower. This is because some diseases (such as myocardial infarction) in MYO lead to changes in its appearance, which increases the difficulty. However, the invented method still achieves good accuracy.
[0072] It is easy for those skilled in the art to understand that the above are only preferred embodiments of the present invention and should not be regarded as limiting the present invention. Any modifications, equivalent substitutions and improvements that are consistent with the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for cardiac magnetic resonance image segmentation based on a dilated convolutional network with directional field features, characterized in that: The model adopts the U-Net model architecture, including encoder, decoder and direction field module, and its steps are: (1) In the decoder and encoder, the two consecutive 3×3 convolutional layers of the original U-Net are replaced by a more efficient dilated convolutional module; (2) Secondly, the model upsamples the feature size of each layer of the decoder to the same size as the original image, and then connects and fuses them through a 1×1 convolutional layer. The fused features will be used as the input of the direction field module; (3) Finally, the model refines the fused features through the direction field to generate the final segmentation result. The output segmentation map contains four channels, representing the probabilities of LV, RV, MYO and background respectively.
2. The method for cardiac magnetic resonance image segmentation based on a dilated convolutional network with directional field features according to claim 1, characterized in that: The model combines a deep learning-based medical image segmentation method and a dilated convolution module into a unified network that generates an initial feature map using an encoder and a decoder; In the encoder, the input image passes through five dilated convolution modules with dilation rates of 1, 2, and 3 respectively. The number of convolution kernels of the five dilated convolutions is 64, 128, 256, 512, and 512 respectively, and the step size is 1. The SAME method is used for padding and the linear rectification function is used as the activation function. The last layer of each module is a 2*2 maximum pooling layer, and the final output is a matrix of size 16*16*512; In the decoder, the input image passes through a bilinear interpolation upsampling layer with a magnification of 2; all upsampled feature maps need to be jump-connected with the output of the same layer in the encoder, so that the feature maps of the same layer are added at the channel level and then input into the dilated convolution module; the decoder contains four dilated convolution modules with dilation rates of 1, 2 and 3 respectively. The number of convolution kernels of the four dilated convolutions is 256, 128, 64 and 64 respectively, and the step size is 1. The SAME method is used for padding and the linear rectification function is used as the activation function, and finally a matrix of size 256*256*64 is output; finally, the features output by the fourth layer of the decoder are aggregated into a 256*256*512 matrix in the channel dimension. This layer uses a convolution kernel of size 1*1 to reduce the number of channels. The number of convolution kernels is 64, the step size is 1, the SAME method is used for padding and the linear rectification function is used as the activation function, and a matrix of size 256*256*64 is output.
3. The method for cardiac magnetic resonance image segmentation based on a dilated convolutional network with directional field features according to claim 1, characterized in that: In order to make the segmentation more accurate, the invention uses a direction field module composed of 1×1 convolution to learn the direction field. Its input is a matrix of size 256*256*64 from the output of the first decoder, and the output is a direction field with 2 channels. The background pixel of the direction field is (0, 0), which is defined as follows: Where a represents the foreground pixel, b represents the pixel closest to the cardiac tissue boundary, ba→ is the direction vector between b and a, and the invention is normalized by distance. The direction field module provides a direction vector for each pixel to point to the central area, thereby predicting the relationship between pixels. After generating the direction field, the model uses the generated direction field F to improve the output feature M 0 To obtain the improved feature M N The features in the center area are error corrected to M 0 ,Each pixel is iteratively updated, and the operation is defined as follows: Where Ω is the image domain, k represents the kth step, N is the total number of iterations, and p x and p y are the x and y coordinates of pixel a respectively. Then, M N and M 0 Concatenate them together and use them as the input of the final classifier to generate the segmentation result.
4. The method for cardiac magnetic resonance image segmentation based on a dilated convolutional network with directional field features according to claim 1, characterized in that: The loss function involved in the invention uses cross entropy LCE as the segmentation loss, and LCE is defined as follows: L CE =-∑p i log2(q i ) where p i is the true value, q i is the predicted value. Then the model selects L2 norm distance and angle distance as the loss of direction field learning: In the formula, F and are the ground truth and the corresponding predicted direction field, respectively. The hyperparameter α is set to 1 to balance the L2-norm distance and the angular distance. The weight on pixel a is represented by w(a), which is defined as: Among them, |C i | is the total number of pixels with label i, N cls is the number of classes, the lower bound of the total loss contains L CE and L F , where the balancing factor λ = 1, the total loss is as follows:
Citation Information
Patent Citations
Image processing using convolutional neural network to track plurality of objects
CN113924597A
Cerebrovascular tracking method based on streamline differential equation
CN115147572A
