A blood vessel image segmentation method, system, electronic device and storage medium

Abstract

The application discloses a blood vessel image segmentation method and system, an electronic device and a storage medium. The scheme extracts and fuses features of a blood vessel image to obtain a second feature map and calculate a first loss function value. The second feature map is discriminated and a second loss function value is calculated. The loss function values are compared with corresponding preset values. If the loss function values do not satisfy the corresponding preset values, the second feature map is updated according to a first preset algorithm, and the loss function value is updated. The updated second feature map is discriminated until the loss function value converges to the corresponding preset value, and the updated second feature map is taken as a segmentation result. The features of the blood vessel image are extracted and fused to enrich the feature information of thin and fragile blood vessels and the feature information of complex topological forms of blood vessels, and the blood vessel segmentation accuracy is improved. The feature map is discriminated by a discriminator to constrain blood vessel segmentation and improve the accuracy of the blood vessel segmentation result. The application can be widely applied to the technical field of medical image processing.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and in particular to a method, system, electronic device and storage medium for segmenting blood vessel images. Background Technology

[0002] Retinal vascular images are closely related to ophthalmology and cardiovascular diseases. Segmenting blood vessels from initial medical images plays a crucial role in disease research, related treatment plans, and navigation. However, blood vessels in retinal images exhibit complex spatial topology, and there are significant differences between vessels of varying thicknesses. Performing precise segmentation of blood vessels in retinal images manually by medical experts is extremely difficult, and the results often vary depending on the expert's experience. With the development of computer technology and artificial intelligence, vascular segmentation models based on computer vision, machine learning, and deep learning have been developed for retinal vascular image segmentation. However, existing methods are insufficient in extracting feature information from thin and fragile blood vessels and have low accuracy in recognizing the complex topological morphology of retinal blood vessels, resulting in low vascular segmentation accuracy. Summary of the Invention

[0003] The purpose of this application is to provide a method, system, electronic device and storage medium for blood vessel image segmentation, which can improve the accuracy of blood vessel segmentation.

[0004] To achieve the above objectives, one aspect of this application provides a blood vessel image segmentation method, the method comprising:

[0005] Obtain the blood vessel image to be segmented, and perform preprocessing operations on the blood vessel image to be segmented to obtain a first image set;

[0006] Feature extraction is performed on the first image set to obtain a first feature map set; feature fusion is performed on the first feature map set to obtain a second feature map, and a first loss function value is calculated;

[0007] A discrimination operation is performed on the second feature map, and a second loss function value is calculated; the first loss function value is compared with a first preset value, and the second loss function value is compared with a second preset value; wherein, the discrimination operation includes a coarse-grained discrimination operation and a fine-grained discrimination operation;

[0008] If the first loss function value does not meet the first preset value, or the second loss function value does not meet the second preset value, the second feature map is updated according to the first preset algorithm, the first loss function value is updated according to the updated second feature map, and the process returns to perform a discrimination operation on the second feature map and calculate the second loss function value until the first loss function value converges to the first preset value and the second loss function value converges to the second preset value, and the second feature map is used as the segmentation result.

[0009] In some embodiments, performing feature extraction on the first image set to obtain a first feature map set specifically includes:

[0010] Perform residual convolution on the first image set to obtain the third feature map set;

[0011] A sampling operation is performed on the third feature map set to obtain a first sampled map set; a residual convolution operation is performed on the third feature map set to obtain a fourth feature map set; wherein, the sampling operation includes an upsampling operation or a downsampling operation;

[0012] A spatial feature fusion operation is performed on the first sampled map set and the fourth feature map set to obtain a fifth feature map set;

[0013] Perform residual convolution or sampling operations on the fifth feature map set to obtain the sixth feature map set; perform spatial feature fusion operations on the sixth feature map set to obtain the first feature map set.

[0014] In some embodiments, the sampling operation on the third feature map set to obtain the first sampled map set specifically includes:

[0015] The third feature map set is subjected to a first convolution operation to obtain a first sub-feature map set; the first sub-feature map set is subjected to batch normalization and random deactivation to obtain a second sub-feature map set; wherein, the first convolution operation includes a regular convolution operation and a transposed convolution operation.

[0016] The second sub-feature map set is processed according to a preset activation function to obtain the third sub-feature map set; the third feature map set and the third sub-feature map set are concatenated to obtain the fourth sub-feature map set.

[0017] The fourth sub-feature map set is processed according to the preset activation function to obtain the first sampling map set.

[0018] In some embodiments, performing a residual convolution operation on the third feature map set to obtain a fourth feature map set specifically includes:

[0019] The third feature map set is convolved to obtain the fifth sub-feature map set; the third feature map set is then convolved and batch normalized to obtain the sixth sub-feature map set.

[0020] The fifth sub-feature map set is subjected to batch normalization and random deactivation to obtain the seventh sub-feature map set; the seventh sub-feature map set is processed according to a preset activation function to obtain the eighth sub-feature map set; the eighth sub-feature map set is subjected to convolution to obtain the ninth sub-feature map set; the ninth sub-feature map set is subjected to batch normalization and random deactivation to obtain the tenth sub-feature map set.

[0021] The sixth and tenth sub-feature sets are concatenated to obtain the eleventh sub-feature set; the eleventh sub-feature set is then processed according to the preset activation function to obtain the fourth feature set.

[0022] In some embodiments, the step of performing spatial feature fusion on the first sampled map set and the fourth feature map set to obtain a fifth feature map set specifically includes:

[0023] The first sampling map set and the fourth feature map set are stacked to obtain the seventh feature map set; spatial information is extracted from the seventh feature map set to obtain the eighth feature map set; depth features are extracted from the seventh feature map set to obtain the ninth feature map set; and a convolution operation is performed on the seventh feature map set to obtain the tenth feature map set.

[0024] The eighth feature set, the ninth feature set, and the tenth feature set are concatenated to obtain the eleventh feature set; the eleventh feature set is then subjected to batch normalization to obtain the fifth feature set.

[0025] In some embodiments, the first loss function value is calculated according to a first preset formula; wherein, the first preset formula is:

[0026]

[0027] in, The first loss function value, To counteract the loss function value, The value of the binary cross-entropy loss function. Let γ be the value of the topological connectivity loss function. The weighting factor, δ is Weighting factors.

[0028] In some embodiments, the discriminative operation on the second feature map and the calculation of the second loss function value specifically include:

[0029] A coarse-grained discriminative convolution operation is performed on the second feature map to obtain a coarse-grained image; a fine-grained discriminative convolution operation is performed on the second feature map to obtain a fine-grained image.

[0030] The coarse-grained loss function value is calculated based on the coarse-grained image and the second feature map, and the fine-grained loss function value is calculated based on the fine-grained image and the second feature map;

[0031] The second loss function value is obtained by calculating based on the coarse-grained loss function value, the fine-grained loss function value, and the second preset formula.

[0032] In some embodiments, the second preset formula includes:

[0033]

[0034] in, The total loss function value of the discriminator. The loss function value of the coarse-grained discriminator. α is the loss function value of the fine-grained discriminator, β is the weighting factor of the coarse-grained discriminator, and β is the weighting factor of the fine-grained discriminator.

[0035] To achieve the above objectives, another aspect of this application provides a blood vessel image segmentation system, the system comprising:

[0036] The first module is used to acquire the blood vessel image to be segmented, and to perform preprocessing operations on the blood vessel image to be segmented to obtain a first image set;

[0037] The second module is used to perform feature extraction on the first image set to obtain a first feature map set; perform feature fusion on the first feature map set to obtain a second feature map; and calculate the first loss function value.

[0038] The third module is used to perform a discrimination operation on the second feature map and calculate a second loss function value; compare the first loss function value with a first preset value, and compare the second loss function value with a second preset value; wherein, the discrimination operation includes a coarse-grained discrimination operation and a fine-grained discrimination operation;

[0039] The fourth module is used to update the second feature map according to the first preset algorithm if the first loss function value does not meet the first preset value or the second loss function value does not meet the second preset value, update the first loss function value according to the updated second feature map, and return to perform a discrimination operation on the second feature map and calculate the second loss function value until the first loss function value converges to the first preset value and the second loss function value converges to the second preset value, and the second feature map is used as the segmentation result.

[0040] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0041] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described above.

[0042] The implementation of this application embodiment has the following beneficial effects: This application embodiment provides a blood vessel image segmentation method, system, electronic device, and storage medium. This scheme acquires a blood vessel image to be segmented, preprocesses the image to obtain a first image set, extracts features from the first image set to obtain a first feature map set, fuses features to obtain a second feature map, calculates a first loss function value, discriminates the second feature map, calculates a second loss function value, compares the first loss function value with a first preset value, compares the second loss function value with a second preset value, and if the loss function value does not meet the corresponding preset value, updates the second feature map according to a first preset algorithm and updates the corresponding loss function. The loss function is evaluated, and the updated second feature map is used for discrimination until the loss function value converges to the corresponding preset value. The updated second feature map is then used as the segmentation result. Preprocessing is performed on the blood vessel image to be segmented to enrich its feature information. Feature extraction and feature fusion are performed on the blood vessel image to obtain spatial features and depth information at different scales, enriching the feature information of thin and fragile blood vessels. Feature fusion further enriches the feature information of the complex topological morphology of retinal blood vessels, thereby improving the accuracy of blood vessel image segmentation. The obtained second feature map is then evaluated to obtain the loss function. The blood vessel image segmentation result is constrained based on the loss function and the preset value to improve the accuracy of blood vessel image segmentation. Attached Figure Description

[0043] Figure 1 This is a schematic flowchart of a blood vessel image segmentation method provided in an embodiment of this application;

[0044] Figure 2 This is a schematic diagram of the implementation model of a blood vessel image segmentation method provided in an embodiment of this application;

[0045] Figure 3 This is a schematic flowchart illustrating the steps of determining the first feature set in a blood vessel image segmentation method provided in this application embodiment;

[0046] Figure 4 This is a schematic diagram of the generator network structure in the implementation model of a blood vessel image segmentation method provided in this application embodiment;

[0047] Figure 5 This is a schematic flowchart illustrating the steps of determining the first sampling image set in a blood vessel image segmentation method provided in this application embodiment;

[0048] Figures 6(a) to 6(b) are schematic diagrams of the sampling module in the implementation model of a blood vessel image segmentation method provided in the embodiments of this application;

[0049] Figure 7 This is a schematic flowchart illustrating the steps of determining the fourth feature set in a blood vessel image segmentation method provided in this application embodiment;

[0050] Figure 8 This is a schematic diagram of the structure of the improved convolution module in the implementation model of a blood vessel image segmentation method provided in this application embodiment;

[0051] Figure 9 This is a schematic flowchart illustrating the steps of determining the fifth feature set in a blood vessel image segmentation method provided in this application embodiment;

[0052] Figure 10 This is a schematic diagram of the spatial feature fusion module in the implementation model of a blood vessel image segmentation method provided in this application embodiment;

[0053] Figure 11 This is a flowchart illustrating the steps of determining the first loss function value in a blood vessel image segmentation method provided in this application embodiment;

[0054] Figures 12(a) to 12(c) are schematic diagrams of the coarse-grained discriminator and the fine-grained discriminator in the implementation model of a blood vessel image segmentation method provided in the embodiments of this application;

[0055] Figure 13 This is a schematic diagram of blood vessel segmentation results on a public dataset in a specific embodiment provided in this application.

[0056] Figure 14 This is a structural block diagram of a blood vessel image segmentation system provided in an embodiment of this application;

[0057] Figure 15This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application;

[0058] Wherein, 1, original retinal vessel image; 2, preprocessing operation; 3, first image set; 4, generator; 5, output feature map; 6, coarse-grained discriminator; 7, fine-grained discriminator; A, original retinal vessel image; B, grayscale retinal vessel image; C, image manually segmented by medical experts; D, segmentation result of U-Net model; E, segmentation result of Unet++ model; F, segmentation result of AttentionU-Net model; G, segmentation result of FR-UNet model; H, segmentation result of the embodiment of this application. Detailed Implementation

[0059] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0060] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0061] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0062] Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the embodiments of this application is for the purpose of describing the embodiments of this application only and is not intended to limit this application.

[0063] In clinical practice and medical research, the morphology of blood vessels can be used for surgical planning and disease research. The morphology of retinal vessels is closely related to ophthalmological and cardiovascular diseases, such as glaucoma, diabetes, and hypertension. Therefore, studying the geometric changes of blood vessels in retinal images is crucial. Under current technological conditions, initial medical images of retinal vessels cannot intuitively and accurately reflect their structural information, requiring manual segmentation by medical experts. However, due to the complex spatial topology and large number of retinal vessels, manual segmentation is difficult, and the accuracy is affected by the experience of medical experts. With the development of computer technology, machine learning and deep learning are widely used for vessel segmentation, but existing vessel segmentation models struggle to accurately identify thin, small vessels in the retina. The complex spatial topology of retinal vessels is difficult to convert into inherent patterns, resulting in low recognition accuracy. Furthermore, existing models tend to overfit existing features to unseen vascular structures, leading to image illusions, weak generalization, and false positives.

[0064] In view of this, embodiments of this application provide a method, system, electronic device, and storage medium for segmenting blood vessel images. This scheme uses a full-resolution generative adversarial network based on adversarial and efficient convolution as the generator. The convolutional units in the generator are designed with a parallel interactive structure, and the feature maps output by adjacent convolutional units are stacked to extract global and local spatial features and depth information from the input retinal blood vessel image, reducing the loss of spatial information. A lightweight dual-domain discriminator network structure is constructed as the discriminator, which performs coarse-grained and fine-grained discrimination on the generator's output, improving the accuracy of the generator's output. The lightweight design reduces the computational load of the data processing. A loss function for the generator is constructed based on adversarial loss, binary cross-entropy loss, and topological connectivity loss. The generator's loss function constrains the generator, improving the accuracy of the generator's output and enhancing the structural integrity and continuity of the output.

[0065] Figure 1 This is an optional flowchart of a blood vessel image segmentation method provided in this application embodiment. Figure 1 The method may include, but is not limited to, steps S101 to S104.

[0066] Step S101: Obtain the blood vessel image to be segmented, and perform preprocessing operations on the blood vessel image to be segmented to obtain the first image set;

[0067] Step S102: Perform feature extraction on the first image set to obtain a first feature map set; perform feature fusion on the first feature map set to obtain a second feature map, and calculate the first loss function value;

[0068] Step S103: Perform a discrimination operation on the second feature map and calculate the second loss function value; compare the first loss function value with the first preset value, and compare the second loss function value with the second preset value; wherein, the discrimination operation includes coarse-grained discrimination operation and fine-grained discrimination operation;

[0069] Step S104: If the first loss function value does not meet the first preset value, or the second loss function value does not meet the second preset value, the second feature map is updated according to the first preset algorithm, the first loss function value is updated according to the updated second feature map, and the process returns to perform a discrimination operation on the second feature map and calculate the second loss function value until the first loss function value converges to the first preset value and the second loss function value converges to the second preset value, and the second feature map is used as the segmentation result.

[0070] Steps S101 to S104, as illustrated in this embodiment, involve acquiring a medical image of the retinal vessels to be segmented, i.e., the original vessel image 1. Preprocessing operations 2, such as grayscale conversion and normalization, are then performed on the image to be segmented to obtain the first input image set 3. This improves the image quality of the vessel image to be segmented, highlights image differences, and reduces the impact of geometric transformations on the image. Then, upsampling, downsampling, and convolution processing are performed on the input image data. In this embodiment, the following steps are set... Figure 2 The Generator 4 network shown implements the above method. This Generator 4 network is trained using manually segmented and labeled retinal vessel images by a professional physician and several original vessel images. A generator loss function is set to adjust the generator parameters, including an adversarial loss function. Binary cross-entropy loss function and topological connectivity loss function The specific loss function is as follows:

[0071]

[0072] Where N is the total number of pixels in the input feature image, y i ′ Labels generated after images manually segmented and labeled by professional physicians are processed by a dual-domain discriminator. The generator generates predicted images and then processes them with a dual-domain discriminator to label them. P O ∈Dgm(O), P L ∈Dgm(L), Let μ be the bidirectional Hausdorff distance, and μ be the distance between P and P. O A pixel in P, where v is P L One of the pixels.

[0073] The generator extracts feature data from the image and performs feature fusion to output the feature. Figure 5Then, the obtained features Figure 5 The input is processed by a pre-trained lightweight dual-domain discriminator, which includes a coarse-grained discriminator 6 and a fine-grained discriminator 7. The lightweight dual-domain discriminator discriminates the input feature map at different granularities, calculates the loss value of the feature map and the discriminator, and compares the loss value with the corresponding preset threshold to determine the realism of the output feature map. If the loss value does not meet the preset threshold, it means that the predicted segmentation map output by the generator differs significantly from the realism and accuracy of the original blood vessel image. The feature map output by the generator is then updated using a preset neighbor adaptive threshold iteration algorithm. The updated feature map is then re-input into the dual-domain discriminator for discrimination, and the loss value is updated again. The updated loss value is then compared with the corresponding preset threshold. This process is repeated until the loss value converges to the corresponding preset threshold. The latest updated feature map is then used as the segmentation result of the blood vessel medical image.

[0074] In step S101 of some embodiments, the original retinal vessel image can be segmented to enrich the image features and thus improve the accuracy of vessel segmentation; or an algorithm can be used to process the original retinal vessel image to improve image clarity and the accuracy of segmentation results, and this is not limited to these methods.

[0075] Please see Figure 3 In some embodiments, step S102 may include, but is not limited to, steps S301 to S304:

[0076] Step S301: Perform residual convolution operation on the first image set to obtain the third feature map set;

[0077] Step S302: Perform a sampling operation on the third feature map set to obtain the first sampled map set; perform a residual convolution operation on the third feature map set to obtain the fourth feature map set; wherein, the sampling operation includes an upsampling operation or a downsampling operation;

[0078] Step S303: Perform spatial feature fusion operation on the first sampling map set and the fourth feature map set to obtain the fifth feature map set;

[0079] Step S304: Perform residual convolution or sampling operations on the fifth feature map set to obtain the sixth feature map set; perform spatial feature fusion operations on the sixth feature map set to obtain the first feature map set.

[0080] In step S301 of some embodiments, feature extraction is performed on the preprocessed original image using the trained convolutional network. In this embodiment, a full-resolution convolutional parallel interactive generator network is used for feature extraction. The structure of the full-resolution convolutional parallel interactive generator network is as follows: Figure 4 As shown, in Figure 4The convolution module shown includes an improved convolution module and a sampling processing module. The improved convolution module performs residual convolution on the preprocessed image set as input to extract spatial and depth feature information from the image set, resulting in several sets of feature maps.

[0081] In step S302 of some embodiments, several sets of feature maps output by the improved convolution module are sampled accordingly, and convolution processing is performed using an improved convolution module with the same number of feature maps. Further spatial and depth feature information is extracted from the feature maps, and the images are expanded at the horizontal and vertical levels to more comprehensively extract feature information from the original blood vessel image and improve the accuracy of blood vessel image segmentation. In this embodiment, the convolution module processes the original blood vessel image and outputs three sets of feature maps. Then, the same number of improved convolution modules are used to process the three sets of feature maps. Simultaneously, downsampling is performed on the first set of feature maps, upsampling and downsampling are performed on the second set of feature maps, and upsampling is performed on the third set of feature maps to extract feature information at different levels. The feature maps obtained from the sampling processing and the corresponding feature maps obtained from the improved convolution operation are used as input to the next module.

[0082] In step S303 of some embodiments, the spatial feature fusion module spatially fuses the feature map set obtained by the convolutional processing and the feature map set obtained by the sampling processing input from the previous module, stacks and connects adjacent feature map sets obtained by convolutional processing and sampling processing, combines the spatial information and depth features of the backbone network and adjacent networks, and outputs a set of feature map sets to improve the accuracy of the feature map sets in describing the original blood vessel image, thereby improving the accuracy of subsequent blood vessel image segmentation.

[0083] In step S304 of some embodiments, the feature map set output by the spatial fusion module is subjected to improved convolution processing or sampling processing to further extract the spatial information and depth features of the backbone network and adjacent networks. The extracted feature map sets are stacked and the multiple sets of feature maps extracted by the multi-branch network are fused into a set of feature maps as input features for the next stage through spatial feature fusion.

[0084] Please see Figure 5 In some embodiments, step S302 may include, but is not limited to, steps S501 to S503:

[0085] Step S501: Perform a first convolution operation on the third feature map set to obtain a first sub-feature map set; perform batch normalization and random deactivation on the first sub-feature map set to obtain a second sub-feature map set; wherein, the first convolution operation includes regular convolution operation and transposed convolution operation.

[0086] Step S502: Process the second sub-feature map set according to the preset activation function to obtain the third sub-feature map set; perform a concatenation operation between the third feature map set and the third sub-feature map set to obtain the fourth sub-feature map set;

[0087] Step S503: Process the fourth sub-feature map set according to the preset activation function to obtain the first sampling map set.

[0088] In step S501 of some embodiments, the feature map set is sampled using the sampling module shown in Figure 6. Figure 6(a) shows the network structure for upsampling, and Figure 6(b) shows the network structure for downsampling. The sampling operation transmits and interacts the feature information between the backbone network and adjacent networks. Through information interaction between different networks, refined contextual semantic information is supplemented, the spatial model fitting ability is enhanced, and the local receptive field of the feature map is increased. In this embodiment, the feature map set for upsampling is processed using transposed convolutional blocks, which doubles the number of feature map channels and halves the spatial size. The feature map set for downsampling is processed using convolutional blocks, which halves the number of feature map channels and doubles the spatial size. The feature map set obtained through sampling is processed by a batch normalization layer, and then the feature map set is randomly deactivated by a Dropout layer. The feature map set after batch normalization and random deactivation is then input into the next layer for processing.

[0089] In step S502 of some embodiments, the feature map set after normalization and random deactivation is nonlinearly processed by a preset activation function to alleviate the gradient explosion problem caused by feature extraction operation; the feature map set processed by the activation function is concatenated with the feature map set sampled from the input as the input of the next layer; the feature information of the feature map is enriched by residual connection to alleviate gradient explosion.

[0090] In step S503 of some embodiments, the feature map set obtained by stitching in step S502 is processed by a preset activation function to complete the acquisition operation of the input feature map set; the gradient explosion problem caused by convolution through activation function is addressed; in this embodiment, the LeakyReLU function is used as the activation function.

[0091] Please see Figure 7 In some embodiments, step S302 may include, but is not limited to, steps S701 to S703:

[0092] Step S701: Perform convolution operation on the third feature map set to obtain the fifth sub-feature map set; perform convolution operation and batch normalization on the third feature map set to obtain the sixth sub-feature map set.

[0093] Step S702: Perform batch normalization and random deactivation on the fifth sub-feature set to obtain the seventh sub-feature set; process the seventh sub-feature set according to the preset activation function to obtain the eighth sub-feature set; perform convolution on the eighth sub-feature set to obtain the ninth sub-feature set; perform batch normalization and random deactivation on the ninth sub-feature set to obtain the tenth sub-feature set.

[0094] Step S703: Concatenate the sixth and tenth sub-feature sets to obtain the eleventh sub-feature set; process the eleventh sub-feature set according to the preset activation function to obtain the fourth feature set.

[0095] In step S701 of some embodiments, an improved convolution module is set in the full-resolution convolutional parallel interactive generator to acquire the input retinal vessel image and extract features from the feature map set. The specific structure of the improved convolution module is as follows: Figure 8 As shown, the feature map set input to the improved convolutional module is processed by convolutional blocks of the backbone path and residual connection path, and then processed by a batch normalization layer as the input to the next layer; spatial and depth feature information is extracted from the feature map through residual blocks, and the gradient problem is alleviated by a batch normalization layer.

[0096] In step S702 of some embodiments, a Dropout layer is set to randomly deactivate the feature map set output by the batch normalization layer on the backbone path to reduce overfitting and improve the computational efficiency of the improved convolutional module. Then, an activation function is used to process the feature map set output by the Dropout layer to increase the non-linear part of the feature map set. The feature map set processed by the activation function is input into a convolutional layer with a kernel size of 3*3 for further processing to extract spatial and depth feature information from the feature map. A batch normalization layer and a Dropout layer are set to perform batch normalization and random deactivation processing on the feature map set output by the convolutional layer as the output of the backbone path.

[0097] In step S703 of some embodiments, the feature map set output by the backbone path and the feature map set output by the residual connection path are concatenated to enrich the spatial and depth feature information of the feature map set; and an activation function is used to process the concatenated feature map set to alleviate the gradient problem in the processing.

[0098] Please see Figure 9 In some embodiments, step S303 may include, but is not limited to, steps S901 to S902:

[0099] Step S901: Stack the first sampling map set and the fourth feature map set to obtain the seventh feature map set; extract spatial information from the seventh feature map set to obtain the eighth feature map set; extract depth features from the seventh feature map set to obtain the ninth feature map set; perform convolution operation on the seventh feature map set to obtain the tenth feature map set.

[0100] Step S902: The eighth, ninth, and tenth feature sets are concatenated to obtain the eleventh feature set; the eleventh feature set is then subjected to batch normalization to obtain the fifth feature set.

[0101] In step S901 of some embodiments, the following is set: Figure 10 The spatial feature fusion module shown connects the feature information of the feature maps extracted by the convolution modules of different paths, providing feature information of different scales to the feature maps to improve the accuracy of blood vessel image segmentation. In order to reduce parameters, the feature maps of adjacent paths are stacked before the spatial feature fusion operation to improve the diversity and effectiveness of spatial feature convolution integration and play a role in feature transmission. Preserving the spatial relationship between features of different scales helps to model and reconstruct coarse and fine information of the image.

[0102] In step S902 of some embodiments, convolutional kernels of different sizes are set to process the stacked feature maps respectively, and then the processed feature maps are spliced ​​to output a set of feature maps that fuse feature information of different scales. In this embodiment, convolutional kernels of size 3*3, size 1*1, and transposed convolutional kernels of size 3*3 are set to process the stacked feature maps respectively to obtain feature maps containing feature information of different scales. The different feature maps are spliced ​​to obtain a set of feature maps that fuse feature information of different scales.

[0103] Please see Figure 11 In some embodiments, step S103 may include, but is not limited to, steps S1101 to S1103:

[0104] Step S1101: Perform a coarse-grained discriminative convolution operation on the second feature map to obtain a coarse-grained image, and perform a fine-grained discriminative convolution operation on the second feature map to obtain a fine-grained image;

[0105] Step S1102: Calculate the coarse-grained loss function value based on the coarse-grained image and the second feature map, and calculate the fine-grained loss function value based on the fine-grained image and the second feature map;

[0106] Step S1103: Calculate the second loss function value based on the coarse-grained loss function value, the fine-grained loss function value, and the second preset formula.

[0107] In step S1101 of some embodiments, the set full-resolution convolutional parallel interactive generator performs a final feature fusion on the feature map set output by the spatial feature fusion module, outputting a single feature map with the same size as the original image input to the full-resolution convolutional parallel interactive generator. The obtained single feature map is then input into the set lightweight dual-domain discriminator for feature extraction at different granularities, resulting in feature maps of different granularities. In this embodiment, a coarse-grained discriminator and a fine-grained discriminator are set. The structure of the discriminator is shown in Figure 12, where Figure 12(a) is a schematic diagram of the coarse-grained discriminator, Figure 12(b) is a schematic diagram of the convolution module in the fine-grained discriminator, and Figure 12(c) is a schematic diagram of the transposed convolution module in the fine-grained discriminator. The coarse-grained discriminator processes the input single feature map through a set 3*3 convolution kernel and processes it through a batch normalization layer. The processed feature map is concatenated with the input single feature map, and the LeakyReLU function is used to process the concatenated feature map to obtain a coarse-grained feature map. The fine-grained discriminator processes the input single feature map through a set 6*6 convolution kernel, and after batch normalization, it is concatenated with the input single feature map. The LeakyReLU function is used to process the concatenated feature map to obtain a first intermediate feature map. The first intermediate feature map is processed through a set 6*6 transposed convolution kernel, and after batch normalization, it is concatenated with the feature map of the input transposed convolution kernel. The LeakyReLU function is used to process the concatenated feature map to output a second intermediate feature map. The first intermediate feature map and the second intermediate feature map are concatenated, and the transposed convolution module shown in Figure 12(c) is used to process the concatenated feature map to obtain a fine-grained feature map.

[0108] In step S1102 of some embodiments, corresponding loss function values ​​are calculated based on the extracted feature maps of different granularities. Different loss function values ​​evaluate the authenticity and accuracy of the feature maps output by the generator from different levels. Different loss function values ​​can also be used to calculate the total loss function value of the dual-domain discriminator to determine whether the feature maps output by the generator can be used as segmentation results. The specific loss function values ​​are as follows.

[0109]

[0110] Where N and M are the total number of pixels in the input image, y i Image labels manually segmented and annotated by professional physicians. Generate predicted image labels for the generator.

[0111] In step S903 of some embodiments, the total loss function value of the discriminator is calculated based on different loss function values ​​to determine whether the feature map output by the generator can be used as the segmentation result of the original input blood vessel image. The specific calculation formula is as follows:

[0112]

[0113] in, The total loss function value of the discriminator. The loss function value of the coarse-grained discriminator. α is the loss function value of the fine-grained discriminator, β is the weighting factor of the coarse-grained discriminator, and β is the weighting factor of the fine-grained discriminator.

[0114] The following is a detailed description and explanation of the solutions in the embodiments of the present invention, using specific application examples:

[0115] Publicly available retinal segmentation datasets from the internet, including the DRIVE dataset, CHASE_DB1 dataset, and STARE dataset, are obtained. These datasets contain a random distribution of several retinal images segmented and annotated by professional physicians. The obtained retinal segmentation datasets are divided into training and testing sets. The training and testing sets are preprocessed, such as through grayscale conversion and normalization. Then, a sliding window technique or random rotation is used to increase the amount of data in the training set. The generator and discriminator of this embodiment are trained using the Adam parameter optimizer. Simultaneously, a neighbor-adaptive threshold iterative algorithm is used to train the discriminator, with the model parameters' weights decaying to 1×10⁻⁶. -5 The initial learning rate is 1×10 -4 The training sample batch size was 256, and the training epochs were set to 40. Cosine annealing was used to reduce the learning rate, and the weight factors α, β, γ, and δ in the model loss function were set to 0.5, 0.5, 100, and 1, respectively. Retinal vessel images to be segmented from publicly available retinal segmentation datasets were input into both the trained model and the existing model for segmentation, resulting in the following: Figure 13 The visualization results shown are as follows: A is the original retinal vessel image, B is the grayscale retinal vessel image, C is the image manually segmented by medical experts, D is the segmentation result of the U-Net model, E is the segmentation result of the Unet++ model, F is the segmentation result of the Attention U-Net model, G is the segmentation result of the FR-UNet model, and H is the segmentation result of the embodiment of this application. Based on the input image to be segmented and the segmentation results, the area under the ROC curve (AUC), accuracy (Acc), sensitivity (Sen), specificity (Spe), F1 score, and the joint ratio of F1 and target range intersection (IOU) are determined, as shown in Tables 1 to 3 below. Table 1

[0116]

[0117] Table 2

[0118]

[0119] Table 3

[0120]

[0121] Continued from Table 3

[0122]

[0123] As shown in Tables 1, 2, and 3, the embodiments of this application demonstrate excellent overall performance on the DRIVE and STARE datasets in the publicly available datasets. For the STARE dataset, the Acc, Sen, AUC, F1, and IOU of the embodiments of this application reach the highest values ​​of 0.9841, 0.9020, 0.9972, 0.8949, and 0.8100, respectively, with Spe only 0.23% lower than the optimal Spe value. For the DRIVE dataset, the AUC of the embodiments of this application reaches the highest value of 0.9891, and the IOU reaches the highest value of 0.7124, with F1 and Acc values ​​also at the optimal levels. For the CHASE_DB1 dataset, the overall performance of the embodiments of this application is AUC of 0.9906, only 0.07% lower than the model with the highest AUC. ACC, Spe, F1, and IOU rank first, demonstrating strong feature extraction capabilities. Furthermore, the embodiments of this application achieve the highest Sen value on the STARE dataset, ensuring the extraction of thin and delicate blood vessel features and improving the integrity of the vascular structure.

[0124] Using the U-Net model as a benchmark, ablation experiments were conducted using the generator in the model provided in this application embodiment. The coarse-grained discriminator, fine-grained discriminator, and a preset neighbor-adaptive threshold iterative algorithm provided in this application embodiment were gradually added to the experiments. The model's generation and prediction results for the DRIVE dataset were compared, and the results are shown in the table below.

[0125] Table 4

[0126]

[0127] As shown in Table 4, compared with the U-Net model, the generator in this embodiment improves the effect of blood vessel segmentation. The coarse-grained discriminator, the fine-grained discriminator, and the neighbor adaptive threshold iteration algorithm all improve the performance of the generator in blood vessel segmentation and increase the accuracy of blood vessel segmentation.

[0128] Cross-experiments were conducted on the embodiments of this application based on publicly available retinal segmentation datasets. 50% of the DRIVE dataset, CHASE_DB1 dataset, and STARE dataset were used as training sets, and the remaining datasets were used as test sets. Based on the vessel segmentation results of the embodiments of this application, the area under the ROC curve (AUC), accuracy (Acc), sensitivity (Sen), specificity (Spe), F1 score, and the joint ratio of F1 score to target range intersection (IOU) were determined, resulting in the table below.

[0129] Table 5

[0130]

[0131] As shown in Table 5, the embodiments of this application exhibit extremely high Acc and Spe when using the DRIVE dataset as the training set, and excellent Spe and AUC when using the STARE dataset and CHASE_DB1 dataset as the training sets, respectively. Therefore, the embodiments of this application have good generalization ability and robustness.

[0132] The implementation of this application embodiment has the following beneficial effects: This application embodiment provides a blood vessel image segmentation method, system, electronic device, and storage medium. This scheme acquires a blood vessel image to be segmented, preprocesses the image to obtain a first image set, extracts features from the first image set to obtain a first feature map set, fuses features to obtain a second feature map, calculates a first loss function value, discriminates the second feature map to obtain a second loss function value, compares the first loss function value with a first preset value, compares the second loss function value with a second preset value, if the loss function value does not meet the corresponding preset value, updates the second feature map according to a first preset algorithm, updates the corresponding loss function value, and discriminates the updated second feature map until the loss function value is reached. The numerical values ​​converge to the corresponding preset values, and the updated second feature map is used as the segmentation result. By preprocessing the blood vessel image to be segmented, the feature information of the blood vessel image to be segmented is enriched. Through feature extraction and feature fusion, spatial features and depth information at different scales in the blood vessel image are obtained, enriching the feature information of thin and fragile blood vessels in the blood vessel image. Furthermore, feature fusion enriches the feature information of the complex topological morphology of retinal blood vessels, thereby improving the accuracy of blood vessel image segmentation. The obtained second feature map is discriminated to obtain a loss function. The blood vessel image segmentation result is constrained based on the loss function and preset values ​​to improve the accuracy of blood vessel image segmentation. By setting the loss function of feature extraction, the connectivity of the extracted feature information is constrained, thereby improving the continuity of the blood vessel segmentation result.

[0133] like Figure 14 As shown in the figure, this application embodiment also provides a blood vessel image segmentation system that can implement the above-mentioned blood vessel image segmentation method. The system includes:

[0134] The first module is used to acquire the blood vessel image to be segmented, and to perform preprocessing operations on the blood vessel image to be segmented to obtain a first image set;

[0135] The second module is used to perform feature extraction on the first image set to obtain a first feature map set; perform feature fusion on the first feature map set to obtain a second feature map; and calculate the first loss function value.

[0136] The third module is used to perform a discrimination operation on the second feature map and calculate a second loss function value; compare the first loss function value with a first preset value, and compare the second loss function value with a second preset value; wherein, the discrimination operation includes a coarse-grained discrimination operation and a fine-grained discrimination operation;

[0137] The fourth module is used to update the second feature map according to the first preset algorithm if the first loss function value does not meet the first preset value or the second loss function value does not meet the second preset value, update the first loss function value according to the updated second feature map, and return to perform a discrimination operation on the second feature map and calculate the second loss function value until the first loss function value converges to the first preset value and the second loss function value converges to the second preset value, and the second feature map is used as the segmentation result.

[0138] It is evident that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0139] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described blood vessel image segmentation method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0140] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0141] Please see Figure 15 , Figure 15 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0142] The processor 1501 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0143] The memory 1502 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1502 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1502 and is called and executed by the processor 1501 using the blood vessel image segmentation method of the embodiments of this application.

[0144] The input / output interface 1503 is used to implement information input and output;

[0145] The communication interface 1504 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0146] Bus 1505 transmits information between various components of the device (e.g., processor 1501, memory 1502, input / output interface 1503, and communication interface 1504);

[0147] The processor 1501, memory 1502, input / output interface 1503 and communication interface 1504 are connected to each other within the device via bus 1505.

[0148] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described blood vessel image segmentation method.

[0149] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0150] The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. The memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include remote memory located remotely relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0151] It is understood that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0152] This application provides a blood vessel image segmentation method, a blood vessel image segmentation system, an electronic device, and a storage medium. The scheme acquires a blood vessel image to be segmented, preprocesses the image to obtain a first image set, extracts features from the first image set to obtain a first feature map set, fuses features to obtain a second feature map, and obtains a first loss function value. The second feature map is then judged to obtain a second loss function value. The first loss function value is compared with a first preset value, and the second loss function value is compared with a second preset value. If the loss function value does not meet the corresponding preset value, the second feature map is updated according to a first preset algorithm. The loss function value is updated, and the updated second feature map is used for discrimination until the loss function value converges to the corresponding preset value. The updated second feature map is then used as the segmentation result. By extracting features from the blood vessel image to be segmented, spatial features and depth information at different scales in the blood vessel image are obtained, enriching the feature information of thin and fragile blood vessels in the blood vessel image. Feature fusion is also used to enrich the feature information of the complex topological morphology of retinal blood vessels, thereby improving the accuracy of blood vessel image segmentation. The obtained second feature map is then used for discrimination, and the loss function is calculated. The blood vessel image segmentation result is constrained based on the loss function and the preset value to improve the accuracy of blood vessel image segmentation.

[0153] The above is a detailed description of the preferred embodiments of this application. However, the invention of this application is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for segmenting blood vessel images, characterized in that, The method includes: Obtain the blood vessel image to be segmented, and perform preprocessing operations on the blood vessel image to be segmented to obtain a first image set; The first image set is subjected to feature extraction to obtain a first feature map set; the first feature map set is subjected to feature fusion to obtain a second feature map, and a first loss function value is calculated; wherein, the first image set is subjected to feature extraction to obtain a first feature map set includes performing a residual convolution operation on the first image set to obtain a third feature map set; A sampling operation is performed on the third feature map set to obtain a first sampled map set; a residual convolution operation is performed on the third feature map set to obtain a fourth feature map set; wherein, the sampling operation includes an upsampling operation or a downsampling operation; A spatial feature fusion operation is performed on the first sampled map set and the fourth feature map set to obtain a fifth feature map set; The fifth feature map set is subjected to residual convolution or sampling operations to obtain a sixth feature map set; the sixth feature map set is subjected to spatial feature fusion operations to obtain a first feature map set; wherein, the sampling operation on the third feature map set to obtain a first sampled map set includes performing a first convolution operation on the third feature map set to obtain a first sub-feature map set; the first sub-feature map set is subjected to batch normalization and random deactivation to obtain a second sub-feature map set; wherein, the first convolution operation includes regular convolution operation and transposed convolution operation; The second sub-feature map set is processed according to a preset activation function to obtain the third sub-feature map set; the third feature map set and the third sub-feature map set are concatenated to obtain the fourth sub-feature map set. The fourth sub-feature map set is processed according to the preset activation function to obtain the first sampling map set; The step of performing residual convolution on the third feature map set to obtain the fourth feature map set includes performing convolution on the third feature map set to obtain the fifth sub-feature map set; and performing convolution and batch normalization on the third feature map set to obtain the sixth sub-feature map set. The fifth sub-feature map set is subjected to batch normalization and random deactivation to obtain the seventh sub-feature map set; the seventh sub-feature map set is processed according to a preset activation function to obtain the eighth sub-feature map set; the eighth sub-feature map set is subjected to convolution to obtain the ninth sub-feature map set; the ninth sub-feature map set is subjected to batch normalization and random deactivation to obtain the tenth sub-feature map set. The sixth and tenth sub-feature sets are concatenated to obtain the eleventh sub-feature set; the eleventh sub-feature set is then processed according to the preset activation function to obtain the fourth feature set. A discrimination operation is performed on the second feature map, and a second loss function value is calculated; the first loss function value is compared with a first preset value, and the second loss function value is compared with a second preset value; wherein, the discrimination operation includes a coarse-grained discrimination operation and a fine-grained discrimination operation; If the first loss function value does not meet the first preset value, or the second loss function value does not meet the second preset value, the second feature map is updated according to the first preset algorithm, the first loss function value is updated according to the updated second feature map, and the process returns to perform a discrimination operation on the second feature map and calculate the second loss function value until the first loss function value converges to the first preset value and the second loss function value converges to the second preset value, and the second feature map is used as the segmentation result.

2. The method according to claim 1, characterized in that, The step of performing spatial feature fusion on the first sampled map set and the fourth feature map set to obtain the fifth feature map set specifically includes: The first sampling map set and the fourth feature map set are stacked to obtain the seventh feature map set; spatial information is extracted from the seventh feature map set to obtain the eighth feature map set; depth features are extracted from the seventh feature map set to obtain the ninth feature map set; and a convolution operation is performed on the seventh feature map set to obtain the tenth feature map set. The eighth feature set, the ninth feature set, and the tenth feature set are concatenated to obtain the eleventh feature set; the eleventh feature set is then subjected to batch normalization to obtain the fifth feature set.

3. The method according to claim 1, characterized in that, The first loss function value is calculated according to a first preset formula; wherein, the first preset formula is: in, The first loss function value, To counteract the loss function value, The value of the binary cross-entropy loss function. This represents the value of the topological connectivity loss function. for Weighting factors for Weighting factors.

4. The method according to claim 1, characterized in that, The step of performing a discrimination operation on the second feature map and calculating the second loss function value specifically includes: A coarse-grained discriminative convolution operation is performed on the second feature map to obtain a coarse-grained image; a fine-grained discriminative convolution operation is performed on the second feature map to obtain a fine-grained image. The coarse-grained loss function value is calculated based on the coarse-grained image and the second feature map, and the fine-grained loss function value is calculated based on the fine-grained image and the second feature map; The second loss function value is obtained by calculating based on the coarse-grained loss function value, the fine-grained loss function value, and the second preset formula.

5. The method according to claim 4, characterized in that, The second preset formula includes: in, The total loss function value of the discriminator. The loss function value of the coarse-grained discriminator. The loss function value of the fine-grained discriminator. The weighting factors for the coarse-grained discriminator. This represents the weighting factor for the fine-grained discriminator.

6. A blood vessel image segmentation system, characterized in that, include: The first module is used to acquire the blood vessel image to be segmented, and to perform preprocessing operations on the blood vessel image to be segmented to obtain a first image set; The second module is used to perform feature extraction on the first image set to obtain a first feature map set; perform feature fusion on the first feature map set to obtain a second feature map, and calculate a first loss function value; wherein, the feature extraction on the first image set to obtain the first feature map set includes performing residual convolution on the first image set to obtain a third feature map set; A sampling operation is performed on the third feature map set to obtain a first sampled map set; a residual convolution operation is performed on the third feature map set to obtain a fourth feature map set; wherein, the sampling operation includes an upsampling operation or a downsampling operation; A spatial feature fusion operation is performed on the first sampled map set and the fourth feature map set to obtain a fifth feature map set; The fifth feature map set is subjected to residual convolution or sampling operations to obtain a sixth feature map set; the sixth feature map set is subjected to spatial feature fusion operations to obtain a first feature map set; wherein, the sampling operation on the third feature map set to obtain a first sampled map set includes performing a first convolution operation on the third feature map set to obtain a first sub-feature map set; the first sub-feature map set is subjected to batch normalization and random deactivation to obtain a second sub-feature map set; wherein, the first convolution operation includes regular convolution operation and transposed convolution operation; The second sub-feature map set is processed according to a preset activation function to obtain the third sub-feature map set; the third feature map set and the third sub-feature map set are concatenated to obtain the fourth sub-feature map set. The fourth sub-feature map set is processed according to the preset activation function to obtain the first sampling map set; The step of performing residual convolution on the third feature map set to obtain the fourth feature map set includes performing convolution on the third feature map set to obtain the fifth sub-feature map set; and performing convolution and batch normalization on the third feature map set to obtain the sixth sub-feature map set. The fifth sub-feature map set is subjected to batch normalization and random deactivation to obtain the seventh sub-feature map set; the seventh sub-feature map set is processed according to a preset activation function to obtain the eighth sub-feature map set; the eighth sub-feature map set is subjected to convolution to obtain the ninth sub-feature map set; the ninth sub-feature map set is subjected to batch normalization and random deactivation to obtain the tenth sub-feature map set. The sixth and tenth sub-feature sets are concatenated to obtain the eleventh sub-feature set; the eleventh sub-feature set is then processed according to the preset activation function to obtain the fourth feature set. The third module is used to perform a discrimination operation on the second feature map and calculate a second loss function value; compare the first loss function value with a first preset value, and compare the second loss function value with a second preset value; wherein, the discrimination operation includes a coarse-grained discrimination operation and a fine-grained discrimination operation; The fourth module is used to update the second feature map according to the first preset algorithm if the first loss function value does not meet the first preset value or the second loss function value does not meet the second preset value, update the first loss function value according to the updated second feature map, and return to perform a discrimination operation on the second feature map and calculate the second loss function value until the first loss function value converges to the first preset value and the second loss function value converges to the second preset value, and the second feature map is used as the segmentation result.

7. An electronic device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method as described in any one of claims 1-5.

8. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the method as described in any one of claims 1-5.

Images