Right Ventricle Segmentation Method and Device Based on Feature Reuse and Multi-Scale Weight Convolution

Abstract

The present invention discloses a right ventricle segmentation method and device based on feature reuse and multi-scale weighted convolution. The method comprises: obtaining a benchmark dataset, performing data preprocessing and data partitioning on the benchmark dataset to obtain a training test set and a test dataset; constructing a right ventricle segmentation network, comprising a feature reuse encoding path, a multi-scale weighted convolution encoding path, and a decoding path. The feature reuse encoding path is used to mine the complex internal features of the right ventricle; the multi-scale weighted convolution encoding path is used to extract the edge features of the right ventricle and aggregate edge features of different scales; and the decoding path is used to perform right ventricle segmentation based on the features of the two encoding paths; the right ventricle segmentation network is trained using the training test set; and each MRI image in the test dataset is input into the trained right ventricle segmentation network to obtain a right ventricle segmentation result. The present invention improves the efficiency of right ventricle segmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a right ventricle segmentation method and device based on feature reuse and multi-scale weighted convolution. Background Art

[0002] With the aging of the population and the acceleration of urbanization, unhealthy lifestyles are becoming increasingly prominent among residents, and the incidence of cardiovascular disease is rising. Therefore, the assessment of ventricular function is of great significance for the identification and diagnosis of cardiac pathologies (such as coronary heart disease and cardiomyopathy).

[0003] At present, magnetic resonance imaging technology provides a favorable technical means for the study and diagnosis of cardiac ventricular function. Cardiac index data such as endocardial volume, ventricular mass and ejection coefficient are usually quantified from magnetic resonance images (MRI) and used to analyze the overall and regional functions to evaluate the ventricular function of the heart. However, as a prerequisite for evaluating cardiac index data, the left ventricle and right ventricle must be accurately segmented. However, compared with the left ventricle, the right ventricle has the problems of fuzzy boundaries, irregular cavities, complex crescent-shaped structures and small targets, which brings great challenges to the accurate segmentation of the right ventricle.

[0004] Researchers have proposed a variety of related right ventricle segmentation methods, such as deformation-based, atlas-based, and prior model-based methods. However, these methods have obvious shortcomings, such as deformation-based methods rely on stable functions, atlas-based methods rely on the quality of registration, and prior model-based methods rely on professional knowledge. In addition, their segmentation robustness and accuracy are low. These traditional segmentation methods perform far worse than clinical experts in clinical cardiac diagnosis. Studies have shown that the main reason why the segmentation performance of these models is difficult to improve is that these models are shallow models and lack sufficient complexity to model the fuzzy boundaries, irregular cavities, and complex crescent-shaped structures of the right ventricle, resulting in unsatisfactory segmentation results.

[0005] In recent years, deep learning-based medical image segmentation methods have made significant progress. Compared with traditional segmentation methods, deep learning models have deeper structures and can capture more complex image features and higher-level semantic information, achieving better segmentation performance. Therefore, many deep learning-based methods have been applied to medical image processing tasks, such as cardiac image segmentation. Among them, fully convolutional neural networks (FCNs) and semantic segmentation networks (U-Nets) are commonly used as the basis for medical image segmentation. The fully convolutional neural network is an end-to-end network that uses convolutional neural networks instead of fully connected layers to achieve pixel-level image classification. The semantic segmentation network is an improvement on the fully convolutional neural network and has slightly better segmentation performance than the fully convolutional neural network.

[0006] However, existing deep learning models, represented by semantic segmentation networks, still lag behind clinical experts in the performance of right ventricle segmentation. The main reasons are as follows: (1) The right ventricle has an irregular cavity and a complex crescent-shaped structure, which results in the complex internal features of the right ventricle not being fully utilized during feature extraction; (2) The right ventricle has a fuzzy boundary, which makes it difficult to utilize the edge features of the right ventricle; (3) In cardiac MRI images, most pixels belong to the background, and only 5.1% of the pixels are part of the right ventricle cavity. Therefore, there is a small target problem, which will lead to the degradation of the segmentation performance of the deep learning model and the inability to accurately segment the right ventricle.

[0007] Therefore, some new network structures have been proposed to address the above-mentioned problems in right ventricular segmentation. For example, non-patent literature 1 proposes a Cardiac-DeepIED network structure, which consists of a convolutional long short-term memory neural network (Conv-LSTM) and a dilated convolution. This combination can not only obtain detailed features, but also has a strong advantage in spatiotemporal context learning; non-patent literature 2 proposes a DBAN network structure, which consists of an expansion block and a discriminator. The expansion block is used to extract and focus multi-scale features, and the confidence probability map generated by the discriminator can guide the segmenter to refine the segmentation results.

[0008] While both of the aforementioned network structures can segment the right ventricle, they also suffer from the following issues: First, the complex crescent-shaped edge information of the right ventricle is not properly considered. Second, image features are not fully utilized or, when utilized, introduce more useless information. These issues hinder the accuracy of these two network structures in achieving accurate segmentation results.

[0009] Therefore, how to improve segmentation accuracy is still a technical problem that needs to be solved in current right ventricle segmentation methods.

[0010] Reference List

[0011] Non-patent literature

[0012] Non-patent document 1: "Cardiac-DeepIED: Automatic Pixel-Level Deep Segmentation for Cardiac Bi-Ventricle Using Improved End-to-End Encoder-Decoder Network", Xiuquan Du et al., IEEE Journal of Translational Engineering in Health andMedicine, 2019-02-24

[0013] Non-Patent Document 2: “DBAN: Adversarial Network With Multi-Scale Features for Cardiac MRI Segmentation,” Xinyu Yang et al., IEEE Journal of Biomedical and Health Informatics, October 2, 2020 Summary of the Invention

[0014] Based on this, it is necessary to provide a right ventricle segmentation method and device based on feature reuse and multi-scale weighted convolution to address the above-mentioned technical problems in the background technology.

[0015] To achieve the above objectives, the present invention provides a right ventricle segmentation method based on feature reuse and multi-scale weighted convolution, comprising:

[0016] Obtain a benchmark dataset and perform data preprocessing and data partitioning on the benchmark dataset to obtain a training test set and a detection dataset; the benchmark dataset includes multiple cardiac MRI images and manual segmentation labels of some cardiac MRI images;

[0017] Constructing a right ventricle segmentation network; the right ventricle segmentation network includes three paths, namely a feature multiplexing encoding path, a multi-scale weighted convolution encoding path, and a decoding path;

[0018] The feature multiplexing encoding path includes at least two feature multiplexing modules and at least one first downsampling module; the feature multiplexing module is used to extract the internal features of the right ventricle; the first downsampling module is used to downsample the features output by the feature multiplexing module and input the downsampled features into the next module;

[0019] The multi-scale weighted convolution encoding path includes at least one multi-scale weighted convolution module, at least one second downsampling module and a basic convolution block; the multi-scale weighted convolution module is used to extract edge features of different scales of the right ventricle and aggregate the edge features of different scales; the downsampling module is used to downsample the features output by the multi-scale weighted convolution module and input the downsampled features to the next module; the basic convolution block is used to perform a convolution operation on the features output by the downsampling module;

[0020] The decoding path includes a feature fusion module, at least one upsampling module, at least one connection module and at least one basic convolution block and an output layer; the feature fusion layer is used to fuse the features output by the feature multiplexing coding path and the multi-scale weighted convolution coding path; the upsampling module is used to upsample the features from the previous module and input the upsampled features to the connection module; the connection module is used to combine the features output by the first downsampling module, the second downsampling module and the upsampling module using a skip connection, and input them to the basic convolution block; the basic convolution block is used to perform a convolution operation on the features output by the connection module; the output layer is used to predict the segmentation result based on the features output by the last basic convolution block to obtain the right ventricle segmentation result;

[0021] Training the right ventricle segmentation network using the training and testing sets;

[0022] Each MRI image in the detection data set is input into the trained right ventricle segmentation network to obtain the corresponding right ventricle segmentation result.

[0023] In addition, the present invention also provides a right ventricle segmentation device based on feature reuse and multi-scale weighted convolution, comprising:

[0024] A data acquisition module is used to obtain a benchmark dataset and perform data preprocessing and data segmentation on the benchmark dataset to obtain a training test set and a detection dataset; the benchmark dataset includes multiple cardiac MRI images and manual segmentation labels of some cardiac MRI images;

[0025] A network acquisition module is used to construct a right ventricle segmentation network; the right ventricle segmentation network includes three paths, namely a feature multiplexing encoding path, a multi-scale weighted convolution encoding path, and a decoding path;

[0026] The feature multiplexing encoding path includes at least two feature multiplexing modules and at least one first downsampling module; the feature multiplexing module is used to extract the internal features of the right ventricle; the first downsampling module is used to downsample the features output by the feature multiplexing module and input the downsampled features into the next module;

[0027] The multi-scale weighted convolution encoding path includes at least one multi-scale weighted convolution module, at least one second downsampling module and a basic convolution block; the multi-scale weighted convolution module is used to extract edge features of different scales of the right ventricle and aggregate the edge features of different scales; the downsampling module is used to downsample the features output by the multi-scale weighted convolution module and input the downsampled features to the next module; the basic convolution block is used to perform a convolution operation on the features output by the downsampling module;

[0028] The decoding path includes a feature fusion module, at least one upsampling module, at least one connection module and at least one basic convolution block and an output layer; the feature fusion layer is used to fuse the features output by the feature multiplexing coding path and the multi-scale weighted convolution coding path; the upsampling module is used to upsample the features from the previous module and input the upsampled features to the connection module; the connection module is used to combine the features output by the first downsampling module, the second downsampling module and the upsampling module using a skip connection, and input them to the basic convolution block; the basic convolution block is used to perform a convolution operation on the features output by the connection module; the output layer is used to predict the segmentation result based on the features output by the last basic convolution block to obtain the right ventricle segmentation result;

[0029] A network training module, configured to train the right ventricle segmentation network using the training and test sets;

[0030] The right ventricle segmentation module is used to input each MRI image in the detection data set into the trained right ventricle segmentation network to obtain the corresponding right ventricle segmentation result.

[0031] The present invention provides a right ventricle segmentation method and device based on feature reuse and multi-scale weighted convolution. After training the right ventricle segmentation network using a training and test set, each MRI image in the test data set is input into the trained right ventricle segmentation network. The features of the MRI image are extracted through the feature reuse coding path in the right ventricle segmentation network, and information close to the original MRI image is reused to fully explore the complex internal features of the right ventricle. At the same time, right ventricle edge information and aggregated multi-scale features are extracted through the multi-scale weighted convolution coding path. The features obtained by downsampling in the two coding paths and the features obtained by upsampling in the decoding path are then fused through the decoding path, and the fused features are used to perform right ventricle segmentation. Compared with existing right ventricle segmentation methods, the right ventricle segmentation method based on feature reuse and multi-scale weighted convolution effectively improves the accuracy of right ventricle segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 Schematic diagram of a flow chart of a right ventricle segmentation method based on feature reuse and multi-scale weighted convolution in one embodiment of the present invention;

[0034] Figure 2 This is a structural diagram of a right ventricle segmentation network in one embodiment of the present invention;

[0035] Figure 3 1 is a structural diagram of a first downsampling module, a second downsampling module, and a basic convolutional block of a right ventricle segmentation network in one embodiment of the present invention;

[0036] Figure 4 This is a structural diagram of a feature reuse module of a right ventricle segmentation network in one embodiment of the present invention;

[0037] Figure 5 2. This is a structural diagram of a multi-scale weighted convolution module of a right ventricle segmentation network according to an embodiment of the present invention;

[0038] Figure 6 A Dice metric interval quantity diagram of right ventricular diastole in one embodiment of the present invention;

[0039] Figure 7 A diagram showing the number of Dice metric intervals during right ventricular systole in one embodiment of the present invention;

[0040] Figure 8Graph showing the segmentation effects of different network models during right ventricular diastole in one embodiment of the present invention;

[0041] Figure 9 Graph showing the segmentation effects of different network models during right ventricular systole in one embodiment of the present invention;

[0042] Figure 10 Schematic diagram of the structure of a right ventricle segmentation device based on feature reuse and multi-scale weighted convolution in one embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0044] like Figure 1 As shown, an embodiment of the present invention provides a right ventricle segmentation method based on feature reuse and multi-scale weighted convolution, which specifically includes the following steps:

[0045] Step S10: Obtain a benchmark data set, and perform data preprocessing and data partitioning on the benchmark data set to obtain a training test set and a detection data set.

[0046] In step S10, the benchmark dataset is a publicly available dataset related to medical images, consisting of multiple cardiac MRI images and manually segmented labels for some of these images, drawn by a cardiologist on the short axis of diastole and systole. Data preprocessing includes resolution unification, standardization, and normalization. The training and test sets contain paired MRI images and their manually segmented labels, and the test dataset contains MRI images for which labels need to be predicted. These MRI images are preprocessed cardiac MRI images.

[0047] Preferably, step S10 includes the following steps:

[0048] Step S101: Select the ACDC dataset provided by MICCAI and mark it as the benchmark dataset.

[0049] In step S101, the ACDC (Automated Cardiac Diagnosis Challenge) dataset provided by MICCAI (Medical Image Computing and Computer Assisted Intervention Society) is selected and marked as a benchmark dataset.

[0050] The ACDC dataset consists of scans from 150 patients, 100 of which were used as a training set and 50 as a test set. Each scan was acquired using two MRI scanners with different magnetic intensities. Manual segmentation labels for the 100 scans in the training set are publicly available. The ACDC dataset labels were manually drawn on the short axis of diastole and systole by two cardiologists with 10 years of experience. Each label includes the myocardium, left ventricle, and right ventricle.

[0051] That is, the benchmark dataset contains 100 cardiac MRI images with manual segmentation labels and 50 cardiac MRI images without manual segmentation labels.

[0052] In step S102, the cardiac MRI images and manually segmented labels in the ACDC dataset are processed for resolution unification by an interpolation algorithm, so as to unify the resolutions of the cardiac MRI images and manually segmented labels to a preset resolution; the preset resolution depends on the resolution allowed by the network input.

[0053] In step S102, the preset resolution is determined according to the resolutions of the input image and the output image of the right ventricle segmentation network. Optionally, the resolutions of the input image and the output image are both 224×224.

[0054] Specifically, for the cardiac MRI images in the ACDC dataset, it is detected whether the resolution of the cardiac MRI images is greater than the preset resolution. If it is greater than the preset resolution, the cardiac MRI images are reduced by the resampling interpolation algorithm (INTER_AREA) to unify the resolution of the cardiac MRI images to the preset resolution; if it is less than the preset resolution, the cardiac MRI images are enlarged by the interpolation algorithm (INTER_CUBIC) to unify the resolution of the cardiac MRI images to the preset resolution.

[0055] Secondly, for the manually segmented labels in the ACDC dataset, their resolution is also unified to the preset resolution.

[0056] It is understandable that by reducing images larger than the preset resolution through the resampling interpolation algorithm (INTER_AREA), the ripple phenomenon in the image can be avoided, while by enlarging images smaller than the preset resolution through the interpolation (INTER_CUBIC) algorithm, a better interpolation effect can be ensured, thereby achieving the unification of the image resolution.

[0057] Step S103 : performing standardization and normalization processing on the cardiac MRI images in the ACDC data set in sequence to obtain MRI images.

[0058] In step S103, the calculation formula for the normalization process can be expressed as:

[0059]

[0060] In formula (1), X s is the cardiac MRI image after normalization; X is the pixel matrix representing the cardiac MRI image; μ is the pixel mean of the cardiac MRI image; adj_stddev is the adjusted standard deviation, which can be expressed as:

[0061]

[0062] In formula (2), max(·) is the maximum function; σ is the standard deviation; and N is the number of pixels in the cardiac MRI image.

[0063] Secondly, the calculation formula for normalization can be expressed as:

[0064]

[0065] In formula (3), x n is the normalized pixel value of the x-th pixel in the cardiac MRI image; is the initial pixel value of the x-th pixel in the cardiac MRI image; max and min are the maximum and minimum pixel values in the cardiac MRI image, respectively.

[0066] That is, for each cardiac MRI image in the ACDC dataset, it is first standardized according to formulas (1) to (2), and then normalized according to formula (3), thereby obtaining an MRI image, i.e., a preprocessed cardiac MRI image.

[0067] Step S104: construct a training and testing set based on the MRI images with manual segmentation labels, and construct a detection data set based on the MRI images without manual segmentation labels.

[0068] That is, after the images in the ACDC dataset are processed with resolution unification, standardization, and normalization, MRI images with manual segmentation labels are first selected from the ACDC dataset to construct a training test set. At this time, the training test set contains paired MRI images and their manual segmentation labels. Then, a detection dataset is constructed based on the remaining MRI images without manual segmentation labels. At this time, the MRI images contained in the detection dataset are all images that need to predict labels.

[0069] Furthermore, in order to fully utilize the limited data to enhance the generalization ability of the right ventricle segmentation network, the ACDC dataset may be enhanced. In this case, step S10 may further include the following steps:

[0070] Step S105 , changing the position, transparency, and angle of the MRI image through affine transformation to expand the ACDC dataset.

[0071] In step S105, for the MRI image in the ACDC dataset, the position, transparency, and angle of the MRI image are changed by affine transformation. For example, the MRI image is randomly rotated according to a preset rotation range, which may be (0, 0.2); the MRI image is horizontally and vertically translated according to a preset translation distance, which is 0.05; the MRI image is perspective transformed according to a preset transformation range, which is (0, 0.05); and the MRI image is filled using a nearest neighbor algorithm.

[0072] Furthermore, the ACDC dataset can be expanded by performing elastic deformation, local stretching, and compression on MRI images.

[0073] It should be noted that the ACDC dataset is pre-processed and then enhanced. Accordingly, step S105 is performed after step S103.

[0074] Step S20, constructing a right ventricle segmentation network; the right ventricle segmentation network includes three paths, namely a feature multiplexing encoding path, a multi-scale weighted convolution encoding path and a decoding path.

[0075] In step S20, the feature multiplexing encoding path includes at least two feature multiplexing modules and at least one first downsampling module; the feature multiplexing module is used to extract the internal features of the right ventricle; the first downsampling module is used to downsample the features output by the feature multiplexing module and input the downsampled features to the next module. As a preferred embodiment, Figure 2 As shown in , the feature multiplexing encoding path consists of four feature multiplexing modules (FW) and three first downsampling modules (DS1); Figure 3 As shown in the figure, the first downsampling module consists of a batch normalization layer (BN), a Relu activation function layer, a convolutional layer (Conv) with a convolution kernel size of 1*1, and an average pooling layer (Average Pooling) with a sliding window of 2×2, which are connected in sequence.

[0076] The multi-scale weighted convolution encoding path includes at least one multi-scale weighted convolution module, at least one second downsampling module and a basic convolution block; wherein the multi-scale weighted convolution module is used to extract edge features of different scales of the right ventricle and aggregate edge features of different scales; the downsampling module is used to downsample the features output by the multi-scale weighted convolution module and input the downsampled features to the next module; the basic convolution block is used to perform a convolution operation on the features output by the downsampling module. As a preferred embodiment, Figure 2 As shown in , the multi-scale weight convolutional coding path consists of three feature multiplexing modules (MsWC), three second downsampling modules (DS2) and a basic convolution block (BaseConv); Figure 3 As shown in the figure, the second downsampling module consists of a convolutional layer with a stride of 2 and a convolution kernel size of 3*3; the basic convolutional block consists of two convolutional layers.

[0077] The decoding path includes a feature fusion module, at least one upsampling module, at least one connection module, at least one basic convolution block and an output layer; the feature fusion layer is used to fuse the features output by the feature reuse coding path and the multi-scale weight convolution coding path; the upsampling module is used to upsample the features from the previous module and input the upsampled features to the connection module; the connection module is used to use a jump connection to combine the features output by the first downsampling module, the second downsampling module and the upsampling module, and input them to the basic convolution block; the basic convolution block is used to perform a convolution operation on the features output by the connection module; the output layer is used to predict the segmentation result based on the features output by the last basic convolution block to obtain the right ventricle segmentation result. As a preferred embodiment, Figure 2 As shown in the figure, the decoding path consists of a feature fusion module (ConcatenationModule), three upsampling modules (Upsampling) with an upsampling window of 2*2, three connection modules (FM+MsWC+BaseConv), three basic convolution blocks and an output layer (Output); the feature fusion module consists of a feature cascade layer (Concatenate) and a convolution layer with a convolution kernel size of 3*3 connected in sequence. The feature fusion module first fuses the features output by the feature reuse encoding path and the multi-scale weight convolution encoding path through the feature cascade layer, and then performs a convolution operation on the features output by the feature cascade layer through the convolution layer and outputs them to the first upsampling module; the output layer consists of a convolution layer with a convolution kernel size of 1*1 and a Sigmoid activation function layer.

[0078] It can be understood that this embodiment is based on the structural framework of the semantic segmentation network, and uses the feature reuse module and the multi-scale weight convolution module as basic building blocks to construct a right ventricle segmentation network. In the right ventricle segmentation network, the parallel feature reuse encoding path and the multi-scale weight convolution encoding path are first used to fully explore the complex internal features of the right ventricle and extract the edge features of the right ventricle and aggregate multi-scale features. The features output by the two encoding paths are then input into the decoding path. Finally, in the decoding path, the features output by the first downsampling module, the second downsampling module and the upsampling module are fused through the connection module, and the features output by the connection module are convolved through the basic convolution block. The convolution operation is then performed on the features output by the connection module, and the convolved features are used to perform right ventricle segmentation. In the decoding path, a convolution layer (i.e., the output layer) with a convolution kernel size of 1*1 and a Sigmoid activation function can be used to predict the right ventricle segmentation result. In the right ventricle segmentation result, each pixel of the input MRI image is classified as the right ventricle area or the background area. Figure 2 In the right ventricle segmentation results, white and black represent the right ventricle area and the background area, respectively.

[0079] Further, if Figure 4 As shown in the figure, the feature reuse module includes two reuse paths, a feature cascade layer, and a random dropout layer. Each reuse path consists of a first convolution unit, a combination unit, a feature cascade layer, and a second convolution unit. The reuse path first performs two convolution operations on the input MRI image through the first convolution unit and the combination unit, then fuses the input MRI image with the feature map after the two convolution operations through the feature cascade layer, and finally filters the fused features through the second convolution unit. The feature cascade layer is used to fuse the features output by the two reuse paths; the random dropout layer (Drouput) is used to randomly drop the features output by the feature cascade layer. Among them, the first convolution unit consists of a convolution layer with a convolution kernel size of 3*3 and a Relu activation function; the combination unit consists of a convolution layer with a convolution kernel size of 3*3, a Relu activation function, a random dropout layer, and a batch normalization layer connected in sequence; the second convolution unit consists of a convolution layer with a convolution kernel size of 3*3 and a Relu activation function, and this convolution layer is equipped with four groups of filters.

[0080] It can be understood that in the feature reuse module, two reuse paths are first used to reuse the features to obtain useful features of the MRI image as much as possible. Specifically, for each reuse path, two convolution operations are first used to obtain the features of the MRI image, and then the input MRI image is fused with the convolved features to achieve feature reuse. Finally, convolution is used to filter the reused features to retain the target features and remove background information.

[0081] The filtered features are then fused to fully preserve the complex internal features of the right ventricle. Finally, a random dropout layer is used to prevent network overfitting. Compared to the existing dense connection module, the feature reuse module of this embodiment enables the network to reuse useful features while eliminating most useless information.

[0082] Further, if Figure 5 As shown in the figure, the multi-scale weighted convolution module contains four weighted dilated convolution blocks of different scales, a feature cascade layer and a third convolution unit; the four weighted dilated convolution blocks are composed of a dilated convolution unit, a Softmax activation function layer and a weighted layer (Weights), and the four dilated convolution units are composed of a dilated convolution layer, a random dropout layer and a Relu activation function; the third convolution unit is composed of a convolution layer with a stride of 2 and a convolution kernel size of 3×3 and a Relu activation function; the convolution kernel size of the four dilated convolution layers is 3*3, and the expansion rates are 1, 2, 4 and 8 respectively.

[0083] It can be understood that in the multi-scale weighted convolution module, the edge features of the right ventricle at different scales are first obtained through 4 weighted hole convolution blocks. Specifically, the input MRI image is first subjected to multi-scale convolution, and 4 convolution layers with different expansion rates (1, 2, 4 and 8) and convolution kernel sizes of 3*3 are applied to extract features of different scales (3*3, 5*5, 9*9 and 17*17); secondly, the Softmax activation function is used to control the value of the feature map between [0,1], so that the right network can decide which part of the input image needs more attention by using the value of the feature map, and extract features from the key parts to obtain important information; next, the value of the feature map is used as a weight to multiply the convolved feature for weighting, so that the contribution value of the right ventricular edge area is increased as much as possible during network training, and the feature response of irrelevant background areas is suppressed.

[0084] The features output by the four weighted atrous convolution blocks are then combined through a feature concatenation layer. Finally, downsampling is performed using a convolution layer with a stride of 2 and a kernel size of 3*3, minimizing information loss. The multi-scale weighted convolution module of this embodiment is used to extract right ventricular edge information and aggregate multi-scale features. Utilizing the crescent edge information in MRI images can significantly improve the accuracy of right ventricular segmentation.

[0085] Step S30: training the right ventricle segmentation network using the training test set.

[0086] In step S30, the training test set is randomly divided into a training data set and a test data set according to a preset allocation ratio. During the training process, a cross entropy loss function can be used to obtain the loss value of the right ventricle segmentation network during training, and an Adam optimizer or a stochastic gradient descent method can be used to optimize the network weight parameters of the right ventricle segmentation network. During the testing process, the segmentation performance of the right ventricle segmentation network can be obtained using a preset segmentation performance evaluation index. Preferably, the segmentation performance evaluation index includes but is not limited to the Dice metric (Dice Score, DSC) and the Hausdorff distance (Hausdorff Distance, HD).

[0087] The Dice metric is used to calculate the spatial overlap rate of two separated regions. Assuming that for a certain MRI image i in the test dataset, the right ventricle automatic segmentation contour extracted from the automatic segmentation result output by the right ventricle segmentation network is S, and the right ventricle manual segmentation contour extracted from the manual segmentation label corresponding to MRI image i is T, the calculation formula of the Dice metric can be expressed as:

[0088]

[0089] In formula (4), Dice(·) is the Dice metric; S∩T is the overlap of the two regions. The Dice metric ranges from 0 to 1. A larger Dice metric value indicates more overlap and more accurate right ventricular segmentation network segmentation.

[0090] The Hausdorff distance is used to calculate the asymmetric distance between two regions when they are most separated. Assuming that point P on the right ventricle automatic segmentation contour S and the corresponding point on the right ventricle manual segmentation contour T is Q, then the minimum distance from point P to point Q is:

[0091] d(P,T)=min Q∈T ||PQ|| (5)

[0092] In formula (5), d(·) is the minimum distance from point P to point Q; ||·|| is the Euclidean norm. Based on formula (5), the Hausdorff distance can be expressed as:

[0093]

[0094] In formula (6), Hausdorff(S,T) is the maximum contour point distance between the right ventricle automatic segmentation contour S and the right ventricle manual segmentation contour T, which can represent the similarity of the two contours. The smaller the Hausdorff distance, the closer the two contours are, and the more accurate the segmentation effect of the right ventricle segmentation network is, and vice versa.

[0095] Preferably, step S30 includes the following steps:

[0096] Step S301, dividing the training test set into a training data set and a test data set;

[0097] Step S302, initializing the network weight parameters of the right ventricle segmentation network;

[0098] Step S303: sending the training data set to the right ventricle segmentation network for training to obtain a trained right ventricle segmentation network, and calculating a loss function value based on the automatic segmentation results output by the right ventricle segmentation network and the manual segmentation labels;

[0099] Step S304: testing the right ventricle segmentation network using a test data set, obtaining the automatic segmentation result output by the right ventricle segmentation network, and detecting whether the segmentation performance evaluation index value calculated based on the automatic segmentation result and the manual segmentation label meets the preset index value;

[0100] Step S305: If satisfied, stop training and output the trained right ventricle segmentation network;

[0101] Step S306: If it is not satisfied, the preset optimizer is called to optimize the network weight parameters according to the loss function value, and the right ventricular segmentation network is retrained until the segmentation performance evaluation index value obtained based on the automatic segmentation result output by the right ventricular segmentation network and the manual segmentation label meets the preset index value, the training is stopped, the optimized network weight parameters are retained, and the trained right ventricular segmentation network is output.

[0102] In this embodiment, the preset distribution ratio of the training and test sets is 8:2; the segmentation performance evaluation indicators include Dice metric and Hausdorff distance; the preset optimizer is Adam optimizer, and its learning rate is 1e -4 , batch_size is 5.

[0103] At this time, after the training test set is divided into a training data set and a test data set according to a preset distribution ratio of 8:2, the training data set is first used to feed the right ventricle segmentation network containing the initial network weight parameters for training, and then the trained right ventricle segmentation network is tested using the test data set to obtain the automatic segmentation result output by the right ventricle segmentation network. Next, the right ventricle automatic segmentation contour and the right ventricle manual segmentation contour are extracted from the automatic segmentation result and its corresponding manual segmentation label to calculate the Dice metric value and the Hausdorff distance value. If it is detected that any value of the Dice metric value and the Hausdorff distance value does not meet the corresponding preset index, the Adam optimizer is called to update the network weight parameters according to the loss function value obtained during the training process, and the process returns to step S303. The training test set is fed into the right ventricle segmentation network containing the updated network weight parameters for retraining until the segmentation performance evaluation index value obtained according to the automatic segmentation result output by the right ventricle segmentation network and the manual segmentation label meets the preset index value. The training is stopped, and the network weight parameter with the smallest loss function value is retained, and the trained right ventricle segmentation network is output.

[0104] Step S40: Input each cardiac MRI image in the detection data set into the trained right ventricle segmentation network to obtain the corresponding right ventricle segmentation result.

[0105] The MRI image for which a label needs to be predicted in the detection data set is input into the right ventricle segmentation network trained in step S30 to automatically segment the right ventricle and obtain the right ventricle segmentation result.

[0106] In summary, the right ventricle segmentation method based on feature reuse and multi-scale weighted convolution provided in this embodiment, after training the right ventricle segmentation network using the training test set, inputs each MRI image in the detection data set into the trained right ventricle segmentation network, extracts the features of the MRI image through the feature reuse coding path in the right ventricle segmentation network, and reuses information close to the original MRI image to fully explore the complex internal features of the right ventricle; at the same time, extracts the right ventricle edge information and aggregates multi-scale features through the multi-scale weighted convolution coding path, and then fuses the features obtained by downsampling in the two coding paths and the features obtained by upsampling in the decoding path through the decoding path, and uses the fused features to perform right ventricle segmentation. Compared with the existing right ventricle segmentation method, the right ventricle segmentation method based on feature reuse and multi-scale weighted convolution improves the accuracy of right ventricle segmentation.

[0107] In an optional embodiment, the M&MS (Multi-Centre, Multi-Vendor & Multi-Disease Cardiac Image Segmentation Challenge) dataset can be selected as the benchmark dataset. The M&MS dataset is preprocessed to unify the resolution of the cardiac MRI images in the M&MS dataset to 224*224, and then normalized. Next, the preprocessed M&MS dataset is used to train a right ventricle segmentation network, and the trained right ventricle segmentation network is used to perform medical image segmentation.

[0108] In an optional embodiment, to verify the effectiveness of each component in the right ventricle segmentation network, three network models with slightly different structures were established in an ablation experiment for performance comparison, as shown in Table 1. To evaluate the segmentation performance of different network models for the right ventricle region in cardiac MRI images, the Dice metric was used to assess the segmentation performance of the network models.

[0109] Table 1 Ablation experiment components

[0110] Semantic Segmentation Network Feature reuse module Multi-scale weighted convolution module Model 1 √ Model 2 √ √ Model 3 √ √ √

[0111] In the ablation experiment, the three network models are first trained using the training and test sets in the ACDC dataset, and the trained network models are used to segment the right ventricular region of the MRI images in the test dataset. Then, the Dice metric of each MRI image is calculated, and the Dice metric interval is counted, such as Figure 6 and Figure 7 As shown, Figure 6 、 Figure 7 Represents the number of Dice metrics of right ventricular diastole (ED) and right ventricular systole (ES) prediction labels in each interval.

[0112] Depend on Figure 6 It can be seen that model 1 has the least number of high-score intervals and the most number of low-score intervals, so model 1 has the worst segmentation effect; compared with model 1, model 2 has a significantly improved number of high-score intervals; compared with model 2, model 3 has a larger number of high-score intervals, so model 3 has the best segmentation effect and the highest segmentation accuracy in the right ventricular diastole. Figure 7 As can be seen, Models 1, 2, and 3 show a stepped pattern in the high-score range, with the segmentation progress increasing in sequence. Therefore, both the feature reuse module and the multi-scale weighted convolution module can improve the segmentation performance of the right ventricle segmentation network.

[0113] In order to observe the segmentation effects of the above three network models more intuitively, the automatic segmentation results of the three network models are visualized, as shown in the figure. Figure 8 and Figure 9 As shown, Figure 8 、 Figure 9 The automatic segmentation results of the three compared network models during right ventricular diastole and right ventricular systole and the manual segmentation results (GT) by medical experts are shown respectively.

[0114] First, from Figure 8 and Figure 9 Overall, whether it is ventricular diastole or right ventricular systole, the predicted labels of model three are closest to the true labels (i.e., manually segmented labels); from the details, before the feature reuse module is stacked, the segmented predicted labels are prone to incompleteness and cannot completely segment the right ventricular area. After stacking the feature reuse module, the incompleteness phenomenon is significantly improved. This is mainly because the feature reuse module can reuse image features, allowing the deep network to use the useful information of the shallow network, fully explore the complex crescent-shaped structure of the right ventricle, and improve the image segmentation effect. However, in Figure 8 and Figure 9 Through observation of model 2, it can be found that the segmentation effect of the network model in the right ventricular diastole period is not well improved. Therefore, the multi-scale weighted convolution module is continued to be stacked, so that the network model can extract the edge information of the right ventricle and aggregate multi-scale features, thereby enabling the network model to achieve better segmentation effect.

[0115] It should be noted that Model 1 of this embodiment is the existing semantic segmentation network, Model 2 is a network that only stacks a feature reuse module on the basis of the semantic segmentation network; Model 3 is a network that simultaneously stacks a feature reuse module and a multi-scale weight convolution module on the basis of the semantic segmentation network, which is equivalent to the right ventricle segmentation network constructed in this application.

[0116] In an optional embodiment, in order to further evaluate the effectiveness of the right ventricular segmentation network in segmenting the right ventricular diastole and right ventricular systole, the performance of the right ventricular segmentation network can be compared with existing medical image segmentation networks, such as U-Net, ResUNet-a, and Attention U-Net, as shown in Table 2.

[0117] Table 2 Comparative experimental table

[0118]

[0119] In Table 2, ResUNet-a was proposed by Foivos I. Diakogiannis et al. in "ResUNet-a: Adeep learning framework for semantic segmentation of remotely sensed data", and Attention U-Net was proposed by Ozan Oktay in "Attention u-net: Learning where to look for the pancreas".

[0120] As shown in Table 2, the right ventricle segmentation network proposed in this example outperforms other models in terms of both the Dice metric and the Hausdorff distance, particularly the Dice metric (a larger Dice metric indicates a more accurate segmentation). Therefore, the right ventricle segmentation network has the potential to achieve clinical expert-level right ventricle segmentation performance, and its segmentation performance is very similar to that of manual segmentation by clinical experts. Compared with existing methods, the right ventricle segmentation network has better segmentation performance.

[0121] In addition, if Figure 10 As shown, an embodiment of the present invention further provides a right ventricle segmentation device based on feature reuse and multi-scale weighted convolution, including a data acquisition module 110, a network acquisition module 120, a network training module 130 and a right ventricle segmentation module 140. The detailed description of each functional module is as follows:

[0122] The data acquisition module 110 is used to obtain a benchmark dataset and perform data preprocessing and data segmentation on the benchmark dataset to obtain a training test set and a detection dataset; the benchmark dataset includes multiple cardiac MRI images and manual segmentation labels of some cardiac MRI images;

[0123] The network acquisition module 120 is used to construct a right ventricle segmentation network; the right ventricle segmentation network includes three paths, namely a feature multiplexing encoding path, a multi-scale weighted convolution encoding path, and a decoding path;

[0124] The feature multiplexing encoding path includes at least two feature multiplexing modules and at least one first downsampling module; the feature multiplexing module is used to extract the internal features of the right ventricle; the first downsampling module is used to downsample the features output by the feature multiplexing module and input the downsampled features into the next module;

[0125] The multi-scale weighted convolutional encoding path includes at least one multi-scale weighted convolutional module, at least one second downsampling module and a basic convolutional block; the multi-scale weighted convolutional module is used to extract edge features of different scales of the right ventricle and aggregate edge features of different scales; the downsampling module is used to downsample the features output by the multi-scale weighted convolutional module and input the downsampled features to the next module; the basic convolutional block is used to perform a convolution operation on the features output by the downsampling module;

[0126] The decoding path includes a feature fusion module, at least one upsampling module, at least one connection module, at least one basic convolution block and an output layer; the feature fusion layer is used to fuse the features output by the feature multiplexing encoding path and the multi-scale weight convolution encoding path; the upsampling module is used to upsample the features from the previous module and input the upsampled features to the connection module; the connection module is used to combine the features output by the first downsampling module, the second downsampling module and the upsampling module using a skip connection, and input them to the basic convolution block; the basic convolution block is used to perform a convolution operation on the features output by the connection module; the output layer is used to predict the segmentation result based on the features output by the last basic convolution block to obtain the right ventricle segmentation result;

[0127] A network training module 130 is used to train a right ventricle segmentation network using a training test set;

[0128] The right ventricle segmentation module 140 is used to input each MRI image in the detection data set into the trained right ventricle segmentation network to obtain the corresponding right ventricle segmentation result.

[0129] Furthermore, the data acquisition module 110 includes the following submodules, and the detailed description of each functional submodule is as follows:

[0130] The dataset selection submodule is used to select the ACDC dataset as the benchmark dataset;

[0131] The resolution unification submodule is used to unify the resolution of cardiac MRI images and manual segmentation labels in the ACDC dataset through an interpolation algorithm, so as to unify the resolution of cardiac MRI images and manual segmentation labels to a preset resolution;

[0132] The standardization and normalization submodule is used to perform standardization and normalization on the cardiac MRI images in the ACDC dataset in sequence to obtain MRI images;

[0133] The dataset partitioning submodule is used to construct a training and testing set based on MRI images with manual segmentation labels, and to construct a detection dataset based on MRI images without manual segmentation labels.

[0134] Furthermore, the network training module 130 includes the following submodules, and the detailed description of each functional submodule is as follows:

[0135] The data set partitioning submodule is used to partition the training and test sets into training data sets and test data sets;

[0136] The network initialization submodule is used to initialize the network weight parameters of the right ventricle segmentation network;

[0137] The preliminary training submodule is used to input the training data set into the right ventricle segmentation network for training, obtain the trained right ventricle segmentation network, and calculate the loss function value based on the automatic segmentation results output by the right ventricle segmentation network and the manual segmentation labels;

[0138] The network testing submodule is used to input the test data set into the trained right ventricle segmentation network for testing, and detect whether the segmentation performance evaluation index value calculated based on the automatic segmentation results output by the right ventricle segmentation network and the manual segmentation labels meets the preset index value;

[0139] The iterative training submodule is used to call the preset optimizer if it is not satisfied. After optimizing the network weight parameters according to the loss function value, the right ventricle segmentation network is retrained until the segmentation performance evaluation index value calculated based on the automatic segmentation results output by the right ventricle segmentation network and the manual segmentation label meets the preset index value. The training is stopped, the optimized network weight parameters are retained, and the trained right ventricle segmentation network is output.

[0140] The apparatus of the above embodiment is used to implement the corresponding method in the above embodiment and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0141] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present invention is limited to these examples. Within the scope of the present invention, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations in different aspects of the embodiments of the present invention as described above, which are not provided in detail for the sake of simplicity.

[0142] The embodiments of the present invention are intended to encompass all such substitutions, modifications, and variations that fall within the broad scope of the present invention. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention should be included in the scope of protection of the present invention.

Claims

1. A right ventricle segmentation method based on feature reuse and multi-scale weighted convolution, characterized in that: include: Obtain a benchmark data set, and perform data preprocessing and data partitioning on the benchmark data set to obtain a training test set and a detection data set; The benchmark dataset includes multiple cardiac MRI images and manual segmentation labels of some cardiac MRI images; Constructing a right ventricle segmentation network; the right ventricle segmentation network includes three paths, namely a feature multiplexing encoding path, a multi-scale weighted convolution encoding path, and a decoding path; Among them, the feature multiplexing encoding path includes at least two feature multiplexing modules and at least one first downsampling module; the feature multiplexing module includes two multiplexing paths, a feature cascade layer and a random drop layer; each of the multiplexing paths is composed of a first convolution unit, a combination unit, a feature cascade layer and a second convolution unit, and the multiplexing path first performs two convolution operations on the input MRI image through the first convolution unit and the combination unit, and then fuses the input MRI image with the feature map after the two convolution operations through the feature cascade layer, and finally filters the fused features through the second convolution unit; the feature cascade layer is used to fuse the features output by the two multiplexing paths; the random drop layer is used to randomly drop the features output by the feature cascade layer; the first downsampling module is used to downsample the features output by the feature multiplexing module and input the downsampled features to the next module; The multi-scale weighted convolution encoding path includes at least one multi-scale weighted convolution module, at least one second downsampling module and a basic convolution block; the multi-scale weighted convolution module includes four weighted hole convolution blocks of different scales, a feature cascade layer and a third convolution unit; the four weighted hole convolution blocks are composed of a hole convolution unit, a Softmax activation function layer and a weighted layer, and the four hole convolution units are composed of a hole convolution layer, a random drop layer and a Relu activation function; the downsampling module is used to downsample the features output by the multi-scale weighted convolution module and input the downsampled features to the next module; the basic convolution block is used to perform a convolution operation on the features output by the downsampling module; The decoding path includes a feature fusion module, at least one upsampling module, at least one connection module and at least one basic convolution block and an output layer; the feature fusion module is used to fuse the features output by the feature multiplexing coding path and the multi-scale weighted convolution coding path; the upsampling module is used to upsample the features from the previous module and input the upsampled features to the connection module; the connection module is used to combine the features output by the first downsampling module, the second downsampling module and the upsampling module using a skip connection, and input them to the basic convolution block; the basic convolution block is used to perform a convolution operation on the features output by the connection module; the output layer is used to predict the segmentation result based on the features output by the last basic convolution block to obtain the right ventricle segmentation result; Training the right ventricle segmentation network using the training and testing sets; Each MRI image in the detection data set is input into the trained right ventricle segmentation network to obtain the corresponding right ventricle segmentation result.

2. The right ventricle segmentation method based on feature reuse and multi-scale weighted convolution according to claim 1, characterized in that: The first downsampling module consists of a batch normalization layer, a ReLU activation function layer, a convolution layer with a convolution kernel size of 1*1, and an average pooling layer with a sliding window of 2×2, which are connected in sequence.

3. The right ventricle segmentation method based on feature reuse and multi-scale weighted convolution according to claim 2, characterized in that: The first convolution unit consists of a convolution layer with a convolution kernel size of 3*3 and a Relu activation function; the combination unit consists of a convolution layer with a convolution kernel size of 3*3, a Relu activation function, a random drop layer and a batch normalization layer connected in sequence; the second convolution unit consists of a convolution layer with a convolution kernel size of 3*3 and a Relu activation function, and four groups of filters are provided in the convolution layer.

4. The right ventricle segmentation method based on feature reuse and multi-scale weighted convolution according to claim 1, characterized in that: The second downsampling module consists of a convolution layer with a step size of 2 and a convolution kernel size of 3*3; the basic convolution block consists of two convolution layers.

5. The right ventricle segmentation method based on feature reuse and multi-scale weighted convolution according to claim 4, characterized in that: The third convolution unit consists of a convolution layer with a step size of 2, a convolution kernel size of 3×3, and a ReLU activation function; the convolution kernel size of the four void convolution layers is 3*3, and the expansion rates are 1, 2, 4 and 8 respectively.

6. The right ventricle segmentation method based on feature reuse and multi-scale weighted convolution according to claim 1, characterized in that: The feature fusion module in the decoding path is composed of a feature cascade layer and a convolution layer connected in sequence. The feature fusion module first fuses the features output by the feature multiplexing encoding path and the multi-scale weight convolution encoding path through the feature cascade layer, and then performs a convolution operation on the features output by the feature cascade layer through the convolution layer and outputs them to the first upsampling module; The output layer consists of a convolution layer with a convolution kernel size of 1*1 and a Sigmoid activation function layer.

7. The right ventricle segmentation method based on feature reuse and multi-scale weighted convolution according to claim 1, characterized in that: The step of obtaining a benchmark data set and performing data preprocessing and data partitioning on the benchmark data set to obtain a training test set and a detection data set includes: The ACDC dataset is selected as the benchmark dataset; Performing resolution unification processing on the cardiac MRI images and the manually segmented labels in the ACDC dataset by an interpolation algorithm, so as to unify the resolutions of the cardiac MRI images and the manually segmented labels to a preset resolution; performing standardization and normalization processing on the cardiac MRI image in the ACDC dataset in sequence to obtain an MRI image; A training and testing set is constructed based on MRI images with manual segmentation labels, and a detection dataset is constructed based on MRI images without manual segmentation labels.

8. The right ventricle segmentation method based on feature reuse and multi-scale weighted convolution according to claim 1, characterized in that: The step of training the right ventricle segmentation network using the training test set includes: Dividing the training test set into a training data set and a test data set; Initializing network weight parameters of the right ventricle segmentation network; Inputting the training data set into the right ventricle segmentation network for training to obtain a trained right ventricle segmentation network, and calculating a loss function value based on an automatic segmentation result output by the right ventricle segmentation network and a manual segmentation label; Inputting the test data set into the trained right ventricle segmentation network for testing, and detecting whether the segmentation performance evaluation index value calculated based on the automatic segmentation result output by the right ventricle segmentation network and the manual segmentation label meets the preset index value; If it is not satisfied, the preset optimizer is called, and after optimizing the network weight parameters according to the loss function value, the right ventricle segmentation network is retrained until the segmentation performance evaluation index value calculated based on the automatic segmentation result output by the right ventricle segmentation network and the manual segmentation label meets the preset index value, the training is stopped, the optimized network weight parameters are retained, and the trained right ventricle segmentation network is output.

9. A right ventricle segmentation device based on feature reuse and multi-scale weighted convolution, characterized in that: include: A data acquisition module is used to obtain a benchmark data set, and perform data preprocessing and data partitioning on the benchmark data set to obtain a training test set and a detection data set; The benchmark dataset includes multiple cardiac MRI images and manual segmentation labels of some cardiac MRI images; A network acquisition module is used to construct a right ventricle segmentation network; the right ventricle segmentation network includes three paths, namely a feature multiplexing encoding path, a multi-scale weighted convolution encoding path, and a decoding path; Among them, the feature multiplexing encoding path includes at least two feature multiplexing modules and at least one first downsampling module; the feature multiplexing module includes two multiplexing paths, a feature cascade layer and a random drop layer; each of the multiplexing paths is composed of a first convolution unit, a combination unit, a feature cascade layer and a second convolution unit, and the multiplexing path first performs two convolution operations on the input MRI image through the first convolution unit and the combination unit, and then fuses the input MRI image with the feature map after the two convolution operations through the feature cascade layer, and finally filters the fused features through the second convolution unit; the feature cascade layer is used to fuse the features output by the two multiplexing paths; the random drop layer is used to randomly drop the features output by the feature cascade layer; the first downsampling module is used to downsample the features output by the feature multiplexing module and input the downsampled features to the next module; The multi-scale weighted convolution encoding path includes at least one multi-scale weighted convolution module, at least one second downsampling module and a basic convolution block; the multi-scale weighted convolution module includes four weighted hole convolution blocks of different scales, a feature cascade layer and a third convolution unit; the four weighted hole convolution blocks are composed of a hole convolution unit, a Softmax activation function layer and a weighted layer, and the four hole convolution units are composed of a hole convolution layer, a random drop layer and a Relu activation function; the downsampling module is used to downsample the features output by the multi-scale weighted convolution module and input the downsampled features to the next module; the basic convolution block is used to perform a convolution operation on the features output by the downsampling module; The decoding path includes a feature fusion module, at least one upsampling module, at least one connection module and at least one basic convolution block and an output layer; the feature fusion module is used to fuse the features output by the feature multiplexing coding path and the multi-scale weighted convolution coding path; the upsampling module is used to upsample the features from the previous module and input the upsampled features to the connection module; the connection module is used to combine the features output by the first downsampling module, the second downsampling module and the upsampling module using a skip connection, and input them to the basic convolution block; the basic convolution block is used to perform a convolution operation on the features output by the connection module; the output layer is used to predict the segmentation result based on the features output by the last basic convolution block to obtain the right ventricle segmentation result; A network training module, configured to train the right ventricle segmentation network using the training and test sets; The right ventricle segmentation module is used to input each MRI image in the detection data set into the trained right ventricle segmentation network to obtain the corresponding right ventricle segmentation result.

10. The right ventricle segmentation device based on feature reuse and multi-scale weighted convolution according to claim 9, characterized in that: The network training module includes: A data set division submodule, configured to divide the training and test sets into a training data set and a test data set; A network initialization submodule, used to initialize the network weight parameters of the right ventricle segmentation network; a preliminary training submodule, configured to input the training data set into the right ventricle segmentation network for training, obtain a trained right ventricle segmentation network, and calculate a loss function value based on the automatic segmentation results and manual segmentation labels output by the right ventricle segmentation network; A network testing submodule is configured to input the test data set into the trained right ventricle segmentation network for testing, and detect whether the segmentation performance evaluation index value calculated based on the automatic segmentation result output by the right ventricle segmentation network and the manual segmentation label meets the preset index value; The iterative training submodule is used to call the preset optimizer if it is not satisfied, optimize the network weight parameters according to the loss function value, and then retrain the right ventricle segmentation network until the segmentation performance evaluation index value calculated based on the automatic segmentation result output by the right ventricle segmentation network and the manual segmentation label meets the preset index value, stop training, retain the optimized network weight parameters, and output the trained right ventricle segmentation network.

Images