Automatic Segmentation Method of Cardiac Medical Images Based on Lossless Processing

Abstract

The present invention discloses an automatic segmentation method for cardiac medical images based on lossless processing, belonging to the field of medical image analysis. The key points of the technical solution of the present invention include: enlarging the resolution of the cardiac medical image to be segmented by filling pixels around the image, inputting it into a rough segmentation network to obtain a rough segmentation result and removing the filling, finding the coordinates of the center point of the heart in the rough segmentation result, making a fixed-size crop on the cardiac medical image to be segmented with the coordinates of the center point of the heart as the center and recording the crop position, inputting the cropped cardiac region medical image into a network with consistent resolution to obtain a cardiac region segmentation result, and filling the result into the same crop position of the initialized segmentation result to obtain the final segmentation result. The present invention can avoid the problems of pixel information loss in the pooling operation in the convolutional neural network and damage to medical image data in the resizing operation, improve the accuracy of the cardiac medical image segmentation task, and facilitate the diagnosis and treatment process of cardiac diseases by doctors.

Description

Technical Field

[0001] The present invention relates to the field of medical image analysis, and particularly to an automatic segmentation method for cardiac medical images based on lossless processing. Background Art

[0002] The heart is an important organ of the human body. In today's society, many people suffer from heart-related diseases. When suffering from heart diseases, the volume, geometric shape, and motion characteristics of the heart will change to a certain extent. Doctors can observe the state of the heart and analyze the type of disease based on the patient's cardiac medical images. Common medical images include CT (Computed Tomography) images, MRI (Magnetic Resonance Imaging) images, ultrasound images, etc. These images have monotonous colors. For doctors to accurately and completely find different cardiac structures on medical images, they need rich experience and patient observation. Long-term work may also lead to situations such as misjudgment. Automatically segmenting multiple cardiac structures on medical images can greatly facilitate the diagnosis and treatment process of doctors, reduce the workload of doctors, and reduce the probability of misdiagnosis. The results of automatic segmentation of cardiac medical images are also beneficial for technologies such as deep learning to automatically predict and diagnose heart diseases, which helps reduce the cost of physical examinations and enables more people to detect heart diseases periodically or even self-help.

[0003] The task that the automatic segmentation method for cardiac medical images needs to complete is: given a cardiac medical image as input, it can automatically classify each pixel in the input image according to cardiac structures and output a segmentation result with the same resolution as the input image. In the segmentation result, the category of each pixel can be one of the following: background, left ventricle, right ventricle, left myocardium, etc.

[0004] Existing automatic segmentation methods for cardiac medical images with relatively good segmentation accuracy usually work based on deep learning convolutional neural networks, that is, using convolutional neural networks to extract image features and perform layer-by-layer abstraction, so as to complete some advanced visual tasks. For example, the invention patent with the publication number CN112017198A discloses a "right ventricle segmentation method and device based on self-attention mechanism multi-scale features". Its segmentation network is a variant of the convolutional neural network, with the core convolutional and pooling operations of the convolutional neural network. It uses dilated convolution in the deep layer of the network to capture information at different scales, and uses this network structure to segment the right ventricle in magnetic resonance images. This method of segmenting cardiac medical images through convolutional neural networks has the following obvious defects:

[0005] (1) Since the convolutional neural network was originally designed for image classification tasks, the core operation of pooling can expand the receptive field, reduce the resolution of the feature map, and thus reduce the computational complexity of the network. However, in image segmentation tasks, the segmentation accuracy is required at the pixel level, and downsampling operations such as pooling will lose a large amount of pixel information. The network can only calculate and predict the lost pixel information based on the learned parameters during the upsampling stage, which is obviously not conducive to the segmentation accuracy.

[0006] (2) Dilated convolution can also expand the receptive field and can, to a certain extent, replace the operation of convolution plus pooling. However, the resolution of the feature map before and after dilated convolution calculation is the same, which will increase the storage and computational complexity of the network compared with the traditional operation of convolution plus pooling. Therefore, dilated convolution is usually applied to a limited number of layers deep in the network to continue to expand the receptive field and avoid the feature map from being too sparse.

[0007] (3) Even for the same type of medical images, the resolutions of different samples are usually inconsistent, while large-scale training of convolutional neural networks requires the resolutions of samples in the same training batch to be consistent. Existing methods usually use the resizing operation based on the interpolation algorithm to unify the resolutions of all samples, and for the samples in actual applications, the segmentation results of the segmentation network need to be resized back to the previous resolution. However, the resizing operation will add, delete, and modify the values of a large number of pixels based on the interpolation algorithm, making the medical images blurred and the object edges in the segmentation results uneven, which has an obvious damaging effect on medical image data. Summary of the Invention

[0008] In order to overcome the defects that the pooling operation in the convolutional neural network will lose pixel information and thus affect the segmentation accuracy, the dilated convolution will increase the computational complexity and storage of the segmentation network, and the resizing operation based on the interpolation algorithm has a damaging effect on medical image data, etc., the present invention provides an automatic segmentation method for cardiac medical images based on lossless processing.

[0009] The technical solution of the present invention is as follows:

[0010] The present invention provides an automatic segmentation method for cardiac medical images based on lossless processing, and the automatic segmentation method for cardiac medical images based on lossless processing includes:

[0011] S1: Obtain the original cardiac medical image to be segmented and perform normalization processing on it;

[0012] S2: Enlarge the resolution of the original cardiac medical image to be segmented to the input resolution of the rough segmentation network by filling pixels around the cardiac medical image, and record the number of pixels filled around.

[0013] S3: Input the enlarged cardiac medical image to be segmented into the rough segmentation network, calculate the rough segmentation result, and delete an equal number of pixels on the periphery of the rough segmentation result according to the number of filled pixels recorded in step S2;

[0014] S4: Find the coordinates of the center point of the heart on the rough segmentation result;

[0015] S5: On the original cardiac medical image to be segmented, with the coordinates of the center point of the heart as the center and the input resolution of the resolution-consistent network as the size of the cropping window, crop out the cardiac region medical image of the original cardiac medical image to be segmented, and record the cropping position;

[0016] S6: Input the cropped cardiac region medical image into the resolution-consistent network to obtain the cardiac region segmentation result;

[0017] S7: Initialize a segmentation result with the same resolution as the original cardiac medical image to be segmented, and the initial value of all pixels is the same as the pixel value representing the background in the cardiac region segmentation result. Then, according to the cropping position recorded in step S5, fill the cardiac region segmentation result into the same position of the initialized segmentation result to obtain the final segmentation result.

[0018] The resolution-consistent network includes:

[0019] The resolution-consistent network consists of an encoding path and a decoding path;

[0020] The encoding path consists of 5 layers, each layer contains 2 dilated convolution modules, and each dilated convolution module consists of a dilated convolution with a convolution kernel size of 3×3 and a dilation coefficient of 2 i-1 , Batch Normalization operation, and LeakyReLU operation, where i is the index of the layer number of the dilated convolution module in the encoding path or decoding path;

[0021] The decoding path consists of 4 layers and is arranged in reverse order. The feature map output by the 5th layer of the encoding path is input into the 4th layer of the decoding path. Each layer of the decoding path first passes the input feature map through 1 dilated convolution module, then concatenates it with the output feature map of the corresponding layer of the encoding path through a Concatenation operation, and then passes it through 2 dilated convolution modules;

[0022] The output feature maps of the 1st, 2nd, and 3rd layers of the decoding path respectively pass through a convolution operation with a convolution kernel size of 1×1 and the number of convolution kernels consistent with the number of segmentation categories, add the 3 convolution calculation results, and then perform a softmax operation to obtain the prediction result;

[0023] Calculate the segmentation result of the resolution-consistent network according to the method described below: The resolution of the segmentation result is consistent with the resolution of the cardiac medical image input to the resolution-consistent network. The value of each pixel in the segmentation result is the index of the channel with the maximum value among all channels at the same pixel position in the prediction result.

[0024] Optionally, in step S4, use the cardiac center point localization algorithm to find the coordinates of the cardiac center point. The cardiac center point localization algorithm includes:

[0025] Input the annotation or rough segmentation result segmentation, and the number of pixels n skipped during traversal;

[0026] Traverse the pixel values of segmentation in the order from top to bottom and from left to right, find the nth and the penultimate nth pixels that do not represent the background, and record the rows where they are located as x0 and x1 respectively;

[0027] Traverse the pixel values of segmentation in the order from left to right and from top to bottom, find the nth and the penultimate nth pixels that do not represent the background, and record the columns where they are located as y0 and y1 respectively;

[0028] Obtain the rough cardiac center point, and its calculation formula is:

[0029] Starting from the x rough row of segmentation, traverse upward and downward respectively, and find the first row where all pixels in the row represent the background, and record them as x2 and x3 respectively;

[0030] Starting from the y rough column of segmentation, traverse left and right respectively, and find the first column where all pixels in the column represent the background, and record them as y2 and y3 respectively;

[0031] Output the coordinates of the cardiac center point, and its calculation formula is:

[0032] Optionally, before step S3, train the rough segmentation network, including:

[0033] Collect cardiac medical images and corresponding annotations;

[0034] By filling pixels around the image, enlarge the resolution of the cardiac medical image and the corresponding annotation to the input resolution of the rough segmentation network;

[0035] Combine the enlarged cardiac medical image and the corresponding annotation with the batch Dice loss function to train the rough segmentation network.

[0036] Optionally, before step S6, train a resolution consistency network, including:

[0037] Collect cardiac medical images and corresponding annotations;

[0038] Find the coordinates of the cardiac center point on the annotation;

[0039] On the annotation and the corresponding cardiac medical image, with the coordinates of the cardiac center point as the center and the input resolution of the resolution consistency network as the size of the cropping window, crop out the cardiac region in the annotation and the cardiac region in the cardiac medical image;

[0040] Train the resolution consistency network by combining the cropped cardiac region medical images and the corresponding cardiac region annotations with the batch Dice loss function.

[0041] Optionally, the batch Dice loss function includes:

[0042] The calculation formula of the batch Dice loss function is:

[0043]

[0044] In the calculation formula of the batch Dice loss function, c is the index of the segmentation category, i and j are respectively the indices of the cardiac medical images in the same training batch and the indices of the pixels in the cardiac medical image. There are I cardiac medical images in one training batch and J pixels in one cardiac medical image. There are C segmentation categories including the background. P is the prediction result of the segmentation network, G is the one-hot encoded annotation, and e is a decimal to avoid the denominator being zero. This formula is related to the ratio of twice the overlapping area of the prediction result of the segmentation network and the annotation to the sum of the two areas.

[0045] Optionally, the input resolution of the rough segmentation network is 288×288, and the network structure is the U-net network structure widely known in the field.

[0046] Optionally, the number of pixels skipped during the traversal of the input in the cardiac center point positioning algorithm is 20.

[0047] Optionally, the number of convolution kernels in all dilation convolution modules in the resolution consistency network is 96, the input resolution of the resolution consistency network is 112×112, the Adam optimizer is used during training, the number of cardiac medical images in one training batch is 10, the initial learning rate is 0.001, and the learning rate becomes 0.99 times the learning rate in the previous iteration round after each iteration round. The resolution consistency network is trained for 200 iteration rounds.

[0048] The beneficial effects of the present invention include:

[0049] (1) The automatic segmentation method of cardiac medical images based on lossless processing provided by the present invention can avoid the resizing operation based on the interpolation algorithm throughout the process, avoid the damage effect on medical image data, and can solve the problem of inconsistent resolutions of different samples. The medical images input into the segmentation network and the annotations as learning targets are made more real and accurate, and the segmentation results are directly generated by the segmentation network without the need to restore the previous resolution through the resizing operation, so the edges of the segmentation targets will be smoother.

[0050] (2) The automatic segmentation method of cardiac medical images based on lossless processing provided by the present invention can significantly reduce the computational amount and storage amount of the segmentation network, make up for the problem of increased computational amount and storage amount caused by dilated convolution to a certain extent, and enable dilated convolution to be deployed into the entire segmentation network. The method can also enable the segmentation network to focus on the segmentation target area and increase the segmentation accuracy.

[0051] (3) The resolution-consistent network removes all pooling operations, and the resolutions of all feature maps remain consistent. Such a regular network structure has the ability to avoid pixel information loss, is conducive to the learning of the deep layer of the network, and makes full use of the deep layer of the network to complete the segmentation task, increasing the segmentation accuracy. Description of the Drawings

[0052] Figure 1 is the flowchart of the automatic segmentation method of cardiac medical images based on lossless processing provided by the present invention;

[0053] Figure 2 is the cardiac magnetic resonance image to be segmented input in the embodiment of the present invention.

[0054] Figure 3 is in the embodiment of the present invention using the present invention for Figure 2 the segmentation map of three structures of background and heart.

[0055] Figure 4 is the structure diagram of the resolution-consistent network provided by the present invention. Detailed Embodiments

[0056] Combined with the following specific embodiments and drawings, the present invention will be further described in detail. The processes, conditions, experimental methods, etc. for implementing the present invention, except for the specifically mentioned content below, are all common knowledge and well-known common sense in the art. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0057]

Embodiment 1

[0058] The present invention provides an automatic segmentation method of cardiac medical images based on lossless processing, as Figure 1As shown, the automatic segmentation method for cardiac medical images based on lossless processing includes:

[0059] S1: Obtain the original cardiac magnetic resonance image to be segmented, as Figure 2 shown, and perform normalization processing on it;

[0060] S2: By filling pixels around the cardiac magnetic resonance image, enlarge the resolution of the original cardiac magnetic resonance image to be segmented to the input resolution of the rough segmentation network, and record the number of pixels filled around;

[0061] S3: Input the enlarged cardiac magnetic resonance image to be segmented into the rough segmentation network, calculate the rough segmentation result, and delete an equal number of pixels around the rough segmentation result according to the number of filled pixels recorded in step S2;

[0062] S4: Find the cardiac center point coordinates on the rough segmentation result;

[0063] S5: On the original cardiac magnetic resonance image to be segmented, with the cardiac center point coordinates as the center and the input resolution of the resolution-consistent network as the size of the cropping window, crop out the cardiac region magnetic resonance image of the original cardiac magnetic resonance image to be segmented, and record the cropping position;

[0064] S6: Input the cropped cardiac region magnetic resonance image into the resolution-consistent network to obtain the cardiac region segmentation result;

[0065] S7: Initialize a segmentation result with the same resolution as the original cardiac magnetic resonance image to be segmented, and the initial value of all pixels is the same as the pixel value representing the background in the cardiac region segmentation result. Then, according to the cropping position recorded in step S5, fill the cardiac region segmentation result into the same position of the initialized segmentation result to obtain the final segmentation result, as Figure 3 shown.

[0066] Specifically, the resolution-consistent network, as Figure 4 shown, includes:

[0067] The resolution-consistent network consists of an encoding path and a decoding path;

[0068] The encoding path consists of 5 layers, and each layer contains 2 dilated convolution modules. Each dilated convolution module consists of a dilated convolution with a convolution kernel size of 3×3 and a dilation coefficient of 2 i-1 and BatchNormalization operation and LeakyReLU operation, where i is the index of the layer number of the dilated convolution module in the encoding path or decoding path;

[0069] The decoding path consists of 4 layers and is arranged in reverse order. The feature map output by the 5th layer of the encoding path is input to the 4th layer of the decoding path. For each layer of the decoding path, the input feature map first passes through 1 of the dilated convolution modules, and then is concatenated with the output feature map of the corresponding layer of the encoding path through a Concatenation operation, and then passes through 2 of the dilated convolution modules;

[0070] For the output feature maps of the 1st, 2nd, and 3rd layers of the decoding path, convolution operations with a convolution kernel size of 1×1 and the number of convolution kernels being the same as the number of segmentation categories are performed respectively. The results of the 3 convolution calculations are added, and then a softmax operation is performed to obtain the prediction result;

[0071] Calculate the segmentation result of the resolution-consistent network according to the following method: The resolution of the segmentation result is the same as the resolution of the cardiac magnetic resonance image input to the resolution-consistent network. The value of each pixel in the segmentation result is the index of the channel with the maximum value among all channels at the same pixel position in the prediction result.

[0072] Specifically, in step S4, the cardiac center point localization algorithm is used to find the coordinates of the cardiac center point. The cardiac center point localization algorithm includes:

[0073] Input the annotation or rough segmentation result segmentation, and the number of pixels n skipped during traversal;

[0074] Traverse the pixel values of segmentation in the order from top to bottom and from left to right, and find the nth and the penultimate nth pixels that do not represent the background. Denote the rows where they are located as x0 and x1 respectively;

[0075] Traverse the pixel values of segmentation in the order from left to right and from top to bottom, and find the nth and the penultimate nth pixels that do not represent the background. Denote the columns where they are located as y0 and y1 respectively;

[0076] Obtain the rough cardiac center point, and its calculation formula is:

[0077] Starting from the x rough th row of segmentation, traverse upward and downward respectively, and find the first row where all pixels in the row represent the background, and denote them as x2 and x3 respectively;

[0078] Starting from the y rough th column of segmentation, traverse left and right respectively, and find the first column where all pixels in the column represent the background, and denote them as y2 and y3 respectively;

[0079] Output the coordinates of the center point of the heart, and its calculation formula is:

[0080] Specifically, before step S3, train the rough segmentation network, including:

[0081] Collect 1500 cardiac magnetic resonance images and corresponding annotations. The resolution of all cardiac magnetic resonance images and their annotations in the length and width directions does not exceed 288. There are four categories in the annotations: background, left ventricle, right ventricle, and left myocardium, and perform normalization processing on the cardiac magnetic resonance images;

[0082] By filling pixels around the image, enlarge the resolution of the cardiac magnetic resonance image and the corresponding annotation to the input resolution of the rough segmentation network;

[0083] Train the rough segmentation network by combining the enlarged cardiac magnetic resonance image and the corresponding annotation with the batch Dice loss function.

[0084] Specifically, before step S6, train the resolution consistency network, including:

[0085] Collect 1500 cardiac magnetic resonance images and corresponding annotations. The resolution of all cardiac magnetic resonance images and their annotations in the length and width directions does not exceed 288. There are four categories in the annotations: background, left ventricle, right ventricle, and left myocardium, and perform normalization processing on the cardiac magnetic resonance images;

[0086] Input the annotation into the cardiac center point localization algorithm to obtain the coordinates of the cardiac center point;

[0087] On the annotation and the corresponding cardiac magnetic resonance image, with the coordinates of the cardiac center point as the center and the input resolution of the resolution consistency network as the size of the cropping window, crop out the cardiac region in the annotation and the cardiac region in the cardiac magnetic resonance image;

[0088] Train the resolution consistency network by combining the cropped cardiac region magnetic resonance image and the corresponding cardiac region annotation with the batch Dice loss function.

[0089] Specifically, the batch Dice loss function includes:

[0090] The calculation formula of the batch Dice loss function is:

[0091]

[0092] In the calculation formula of the batch Dice loss function, c is the index of the segmentation category, i and j are the indices of the cardiac magnetic resonance images in the same training batch and the indices of the pixels in the cardiac magnetic resonance image respectively. There are I cardiac magnetic resonance images in a training batch, and J pixels in a cardiac magnetic resonance image. There are C segmentation categories including the background. P is the prediction result of the segmentation network, G is the one-hot encoded annotation, and e is a decimal to avoid the denominator being zero. This formula is related to the ratio of twice the overlapping area of the prediction result of the segmentation network and the annotation to the sum of their areas.

[0093] Specifically, the input resolution of the rough segmentation network is 288×288, and the network structure is the U-net network structure commonly known in the field.

[0094] Specifically, the number of pixels skipped during the traversal of the input in the cardiac center point localization algorithm is 20.

[0095] Specifically, the number of convolution kernels in all dilated convolution modules of the resolution consistency network is 96. The input resolution of the resolution consistency network is 112×112. The Adam optimizer is used during training. The number of cardiac magnetic resonance images in a training batch is 10, the initial learning rate is 0.001, and the learning rate becomes 0.99 times the learning rate in the previous iteration round after each iteration round. The resolution consistency network is trained for 200 iteration rounds.

[0096] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any omission, modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An automatic segmentation method for cardiac medical images based on lossless processing, characterized in that, The automatic segmentation method for cardiac medical images based on lossless processing includes: S1: Obtain the original cardiac medical image to be segmented and perform normalization processing on it; S2: Enlarge the resolution of the original cardiac medical image to be segmented to the input resolution of the rough segmentation network by filling pixels around the cardiac medical image, and record the number of pixels filled around; S3: Input the enlarged cardiac medical image to be segmented into the rough segmentation network, calculate the rough segmentation result, and delete an equal number of pixels around the rough segmentation result according to the number of filled pixels recorded in step S2; S4: Find the cardiac center point coordinates on the rough segmentation result; S5: On the original cardiac medical image to be segmented, take the cardiac center point coordinates as the center and the input resolution of the resolution-consistent network as the size of the cropping window, crop out the cardiac region medical image of the original cardiac medical image to be segmented, and record the cropping position; S6: Input the cropped cardiac region medical image into the resolution-consistent network to obtain the cardiac region segmentation result; S7: Initialize a segmentation result with the same resolution as the original cardiac medical image to be segmented, and the initial value of all pixels is the same as the pixel value representing the background in the cardiac region segmentation result. According to the cropping position recorded in step S5, fill the cardiac region segmentation result into the same position of the initialized segmentation result to obtain the final segmentation result.

2. The automatic segmentation method of cardiac medical images based on lossless processing according to claim 1, wherein The resolution-consistent network includes: The resolution-consistent network consists of an encoding path and a decoding path; The encoding path consists of 5 layers, each layer contains 2 dilated convolution modules, and each dilated convolution module consists of a dilated convolution with a convolution kernel size of 3×3 and a dilation coefficient of 2 i-1 , a BatchNormalization operation, and a LeakyReLU operation, where i is the index of the layer number of the dilated convolution module in the encoding path or the decoding path; The decoding path consists of 4 layers and is arranged in reverse order. The feature map output by the 5th layer of the encoding path is input into the 4th layer of the decoding path. Each layer of the decoding path first passes the input feature map through 1 dilation convolution module, then concatenates it with the output feature map of the corresponding layer of the encoding path through the Concatenation operation, and then passes through 2 dilation convolution modules; The output feature maps of the 1st, 2nd, and 3rd layers of the decoding path respectively pass through convolution operations with a convolution kernel size of 1×1 and the number of convolution kernels being the same as the number of segmentation categories, add the 3 convolution calculation results, and then perform the softmax operation to obtain the prediction result; Calculate the segmentation result of the resolution-consistent network according to the following method: The resolution of the segmentation result is the same as the resolution of the cardiac medical image input into the resolution-consistent network. The value of each pixel in the segmentation result is the index of the channel with the maximum value among all channels at the same pixel position in the prediction result.

3. The automatic segmentation method of cardiac medical images based on lossless processing according to claim 1, characterized in that, In step S4, use the cardiac center point localization algorithm to find the cardiac center point coordinates. The cardiac center point localization algorithm includes: Input the labeled or rough segmentation result segmentation and the number of pixels to skip during traversal n; Traverse the pixel values of segmentation in the order from top to bottom and from left to right, find the nth and the penultimate nth pixels that do not represent the background, and record the rows where they are located as x0 and x1 respectively; Traverse the pixel values of the segmentation in the order from left to right and from top to bottom, find the nth and the nth from the bottom non-background pixels, and record the columns where they are located as y0 and y1 respectively; Obtain the rough coordinates of the heart center point, and its calculation formula is: Starting from the x-th rough row of the segmentation, traverse upward and downward respectively to find the first row in which all pixels represent the background, denoted as x2 and x3 respectively; Starting from the y-th column of the segmentation rough Traverse left and right respectively, and find the first columns where all pixels in the column represent the background, denoted as y2 and y3 respectively; Output the coordinates of the center point of the heart, and its calculation formula is:

4. The automatic segmentation method of cardiac medical images based on lossless processing according to claim 1, wherein Before step S3, train the rough segmentation network, including: Collect cardiac medical images and corresponding annotations; By filling pixels around the cardiac medical image, enlarge the resolution of the cardiac medical image and the corresponding annotation to the input resolution of the rough segmentation network; Train the rough segmentation network by combining the enlarged cardiac medical image and the corresponding annotation with the batch Dice loss function.

5. The automatic segmentation method of cardiac medical images based on lossless processing according to claim 1, characterized in that Before step S6, train the resolution consistency network, including: Collect cardiac medical images and corresponding annotations; Find the coordinates of the center of the heart on the annotation; On the annotation and the corresponding cardiac medical image, with the coordinates of the center of the heart as the center and the input resolution of the resolution consistency network as the size of the cropping window, crop out the cardiac region in the annotation and the cardiac region in the cardiac medical image; Train the resolution consistency network by combining the cropped cardiac region medical image and the corresponding cardiac region annotation with the batch Dice loss function.

6. The automatic segmentation method of cardiac medical images based on lossless processing according to claim 4 or 5, characterized in that The batch Dice loss function includes: The calculation formula of the batch Dice loss function is: In the calculation formula of the batch Dice loss function, c is the index of the segmentation category, i and j are respectively the index of the cardiac medical image in the same training batch and the index of the pixel in the cardiac medical image. There are I cardiac medical images in a training batch and J pixels in a cardiac medical image. There are C segmentation categories including the background. P is the prediction result of the segmentation network, G is the one-hot code encoding of the annotation, e is a decimal to avoid the denominator being 0. This formula is related to the ratio of twice the overlapping area of the prediction result of the segmentation network and the annotation to the sum of the two areas.

Images