A contextual learning method for medical image segmentation based on focused fusion

Abstract

The present invention discloses a contextual learning medical image segmentation method based on focus fusion. The method comprises the following steps: preprocessing the original image data; inputting the preprocessed image data into a trained convolutional neural network for feature extraction, wherein the encoding network of the convolutional neural network comprises a plurality of encoding stages and each encoding stage is sequentially connected to an automatic focusing module and a panoramic aggregation module, and the decoding network of the convolutional neural network sequentially decodes the features of the encoding stage and generates corresponding decoding features, wherein the automatic focusing module is used for multi-scale contextual information mining of the same-level encoding features, and the panoramic aggregation module is used for aggregating cross-level encoding features from different encoding stages. The present invention can effectively encode the same-level features and realize the efficient fusion of cross-level features, thereby significantly improving the accuracy of image segmentation.

Description

Technical Field

[0001] The present invention relates to the field of biomedical engineering technology, and more specifically, to a context learning medical image segmentation method based on focus fusion. Background Art

[0002] With the rapid development and popularization of medical imaging equipment, imaging technologies including magnetic resonance imaging (MR), computed tomography (CT), ultrasound, positron emission tomography (PET), etc., have become indispensable equipment for medical institutions to carry out disease diagnosis, surgical planning, prognosis assessment, and follow-up. Medical image processing is the first step in analyzing medical images, which helps to make images more intuitive and clear and improve diagnostic efficiency. Therefore, medical image processing is highly valued both at home and abroad. Image segmentation is an important part of image processing and also a difficult point. Accurate disease detection can provide important and reliable reference value for medical diagnosis. The lesion areas in medical images often show high intra-class differences and unclear distinction between classes. At the same time, some noise will also interfere with image quality. Therefore, how to extract effective context information and how to aggregate different levels of discriminant features are still a great challenge for fine segmentation of images.

[0003] At present, with the increase in the number of medical images and the development of artificial intelligence (AI) technology, computer-aided diagnosis has become an important means. However, it is still a challenging task to automatically identify infected areas in medical images and detect shadows and lesions in images. Among them, high inter-class similarity and intra-class differences are the main problems affecting recognition accuracy. For example, for lung CT images, lung infection lesions have complex characteristic manifestations such as pulmonary consolidation, ground-glass opacity, bilateral patchy shadows and other abnormalities, and the symptoms of infection in patients at different stages are also significantly different. In addition, low contrast and blurred boundaries further increase the difficulty of automatic image segmentation.

[0004] Recently, thanks to the development of Convolutional Neural Network (CNN), many traditional medical image processing problems have made significant breakthroughs. Compared with traditional artificial features, deep learning methods are more effective in extracting high-level semantic features from images. At present, most medical image semantic segmentation networks follow the encoding and decoding framework. Through this structure, the network can mine rich semantic information for target segmentation tasks. For example, U-Net uses skip connections to bridge the same-level information between the encoding layer and the decoding layer, effectively supplementing the information loss in the encoding and decoding process, and achieving excellent performance in the decoding process. U-Net++ uses a mesh-like skip connection structure to integrate semantic levels with different functions, providing better segmentation performance. In order to expand feature representation and retain accurate structural information, conditional random field (CRF), Markov random field (MRF), spatial pyramid pooling and other technologies are adopted. For example, PSPNet uses multiple pyramid pooling in the encoding network to obtain multi-scale context information.

[0005] Generally, the encoding network extracts high-level semantic feature representations from the original input image through continuous multiple convolution operations. The decoding network uses a progressive upsampling (or deconvolution) strategy on the encoded features to obtain the final segmentation result. Cascaded convolution operations can effectively extract high-level semantic information of objects during the encoding process, but at the same time they can also lead to inaccurate object boundary segmentation. Especially for small lesions, many CNN-based models fail to retain sufficient spatial details and long-range contextual information due to repeated pooling operations within CNN. In addition, at each stage, for the encoding subnetwork, the model needs to respond to the intra-class differences caused by the different sizes of infected lesions. CNN networks with a single-scale local receptive field are difficult to adapt to such segmentation tasks with large lesion differences.

[0006] In summary, the main defects of the prior art are mainly:

[0007] 1) Traditional analysis methods, when dealing with medical image segmentation problems, usually perform analysis based on feature points, edge detection, etc., but their disadvantages are that they are easily disturbed by imaging contrast, noise, and some artificial effects, and have low analysis accuracy. In addition, traditional methods often strictly rely on designed artificial features for image encoding, and the prediction performance of complex scenes is usually greatly limited.

[0008] 2) Due to the continuous repetition of pooling operations within CNN, many CNN-based models fail to retain sufficient spatial details and long-range contextual information.

[0009] 3) Since lesions in medical images usually show high intra-class differences and inter-class similarities, the model needs to respond to the intra-class differences caused by the different sizes of infected lesions. CNN networks with a single-scale local receptive field are difficult to adapt to such segmentation tasks with large lesion differences. Summary of the invention

[0010] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a contextual learning medical image segmentation method based on focus fusion, which effectively encodes the same-level features through the automatic focusing module and realizes the efficient fusion of cross-level features through the panoramic aggregation module.

[0011] The technical solution of the present invention is to provide a context learning medical image segmentation method based on focus fusion. The method comprises the following steps:

[0012] Preprocessing the original image data;

[0013] The preprocessed image data is input into a trained convolutional neural network for feature extraction, wherein the encoding network of the convolutional neural network includes multiple encoding stages and each encoding stage is sequentially connected to an automatic focusing module and a panoramic aggregation module, and the decoding network of the convolutional neural network sequentially decodes the features of the encoding stages and generates corresponding decoding features, wherein the automatic focusing module is used for multi-scale contextual information mining of the same-level encoding features, and the panoramic aggregation module is used to aggregate cross-level encoding features from different encoding stages.

[0014] Compared with the prior art, the advantages of the present invention are that fine details and semantic knowledge are effectively extracted through the automatic focusing and panoramic aggregation modules, and long-range contextual dependencies at the same and cross levels are captured; in addition, the new structural consistency correction method is used to effectively improve the segmentation effect of the foreground and background. The present invention can extract rich contextual information and fine-grained features at each stage of the network coding process, and the structural consistency correction further corrects and refines the lesion boundary, significantly improving the accuracy of image segmentation.

[0015] Further features and advantages of the present invention will become apparent from the following detailed description of exemplary embodiments of the present invention with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments of the invention and, together with the description, serve to explain the principles of the invention.

[0017] Figure 1 is a flowchart of a medical image segmentation method based on contextual learning of focus fusion according to an embodiment of the present invention;

[0018] Figure 2 It is an overall framework diagram of a contextual learning medical image segmentation method based on focus fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present invention unless otherwise specifically stated.

[0020] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the invention, its application, or uses.

[0021] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered as part of the specification.

[0022] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0023] It should be noted that like reference numerals and letters refer to similar items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0024] The present invention designs a new convolutional network framework that effectively integrates the contextual information of the same level and across levels of the image. The main body of the model includes an encoding network and a decoding network. In short, in order to obtain a richer feature representation at each stage of the encoding network, an automatic focusing module is designed for multi-scale contextual information mining of features at the same level. Furthermore, a panoramic aggregation module is designed to adaptively aggregate cross-level coding features from different stages of the encoding network, effectively avoiding interference caused by large resolution differences in feature fusion, and further supplementing details and semantic information to the coding features of each stage through cross-layer information complementarity. In addition, a new structural consistency correction method is proposed to supplement the structural level information of the image region in the supervised learning process by constraining the structural relationship between the foreground and background.

[0025] Specifically, combined Figure 1 and Figure 2 As shown, the provided contextual learning medical image segmentation method based on focus fusion includes the following steps:

[0026] Step S110, pre-processing the original image data.

[0027] First, the original image is preprocessed, such as normalization, random flipping, random cropping and other data enhancement methods, to expand and enhance the samples.

[0028] Step S120, constructing a multi-layer convolutional neural network, the network backbone of which includes an encoding network and a decoding network, wherein the encoding network includes multiple encoding stages and each encoding stage is sequentially connected to an automatic focusing module and a panoramic aggregation module to mine multi-scale contextual information of the same-level encoding features and aggregate cross-level encoding features from different encoding stages.

[0029] Combination Figure 2 As shown in Figure 1, a multi-layer convolutional neural network is constructed for feature extraction. For example, the encoding network uses the residual neural network Resnet as the main structure, which contains four residual convolution modules. In the encoding network, the last global pooling layer and the fully connected layer in Resnet are removed, leaving only one convolution and four residual modules for the main feature extraction. Without loss of generality, for each input image, the outputs of the four residual modules are represented as f i , i∈{1,2,3,4}. In order to respond to the diversity of infected lesions and effectively integrate the features of different encoding stages to achieve feature information complementarity, it is preferred that an autofocus module is used at each stage of the encoding network to fully exploit the multi-level features of the same level, and a panoramic aggregation module is designed to aggregate features of different levels. Finally, the accuracy of segmentation is further improved through structural consistency correction. For the sake of clarity, the following will take the encoding network containing four residual convolution modules as an example for specific description.

[0030] 1) About the auto focus module

[0031] The traditional single fixed-size local receptive field is not sufficient to respond to the diversity of lesions. Different convolutional receptive fields have different responses to internal lesions. For example, a large receptive field is suitable for capturing more macroscopic features and describing the relationship between lesions, while a small receptive field is more inclined to extract local manifestations of lesions. In addition, the semantic representations of features generated at different spatial resolutions can effectively correct and complement each other. Therefore, the present invention introduces an automatic focusing mechanism, which constructs a bidirectional continuous receptive field expansion and reduction path, and uses a gating mechanism to aggregate bidirectional contextual information to form a robust feature representation at the current stage.

[0032] First, define a series of transformation operations (such as pooling and convolution), so that the receptive field is gradually expanded (or reduced), and the encoding feature f of each stage is i Perform these transformations respectively to obtain the corresponding feature maps, expressed as:

[0033]

[0034] in, Indicates the corresponding parameters, represents f i The feature map obtained after the tth transformation.

[0035] Furthermore, in order to automatically aggregate the multi-level features in these same-level maps, a novel bidirectional gated context aggregation (bi-GCA) method is proposed to implement the process, which mainly includes a context controller Responsible for synthesizing incoming information and a memory Used to update and store previous information flow, the calculation formula is as follows:

[0036]

[0037] in and Represents three gates, which are used to adaptively select the flow information passing through, σ represents the sigmoid activation function, represents pixel-by-pixel multiplication, represents a 1x1 convolutional layer, represents the corresponding convolution parameters, Indicates the features after the t-th transformation in order The output feature map after integration is Represents the splicing process. Through this aggregation method, the bidirectional aggregation process can be realized as follows:

[0038]

[0039] in Represents the forward integration process The output features of Reverse integration process The output features, F i,AM Represents the feature output after bidirectional integration at the same level of the autofocus module.

[0040] 2) About the Panoramic Aggregation Module

[0041] Due to the continuous convolution and sampling operations, many CNN-based models suffer from the limitation of short-range dependencies of context and cannot simultaneously maintain fine-grained details and extract rich high-level semantic information. In one embodiment, a panoramic aggregation method guided by global information is designed. From a global perspective, a cross-level context integration unit is used to aggregate the Efficiently aggregate encoding information from different stages.

[0042] First, a global context information representation Z is constructed by concatenating and aggregating the feature maps obtained by the autofocus module in each encoding stage, which can be expressed as:

[0043]

[0044] in, represents a 1x1 convolutional layer, represents the splicing process, θ Z Then, the attention mechanism is used to guide this global information into the encoding features of each stage, and the information of each stage is supplemented by the global information:

[0045]

[0046] in, represents the self-attention mechanism, AvgPool represents the global average pooling operation, and σ represents the sigmoid activation function. Through the above steps, the encoding features of each stage are enhanced with global information, θ i,z and Represent the corresponding parameters respectively.

[0047] Next, further integrate the encoding features of different stages into the current stage:

[0048]

[0049] Among them, r i Indicates that the gate function is used to selectively control the transmission of information. represents pixel-by-pixel multiplication, represents the cross-level feature aggregation unit, F i,PM represents the output feature map of the panorama aggregation module at the i-th stage.

[0050] Therefore, features from different stages are integrated through the panoramic aggregation module, and too much redundant information can be avoided. Thanks to the panoramic aggregation module, cross-level context information is effectively integrated and complementary, and the high-level and low-level feature maps of the encoding subnetwork are enhanced.

[0051] Step S130, the decoding network decodes the features of the encoding phase and generates corresponding decoding features.

[0052] Usually, the encoded representation is directly converted into a pixel-level binary segmentation image by using pixel-level constraints for supervision. However, this approach only considers the relationship between each individual pixel and its corresponding true value. The correlation between the surrounding areas of the pixel is ignored, so it is easy to produce inaccurate segmentation results.

[0053] In a preferred embodiment, a new structural consistency correction method is proposed. First, an initial segmentation result is generated. The decoding network decodes the features F of the encoding stage sequentially. i，PM And generate the corresponding decoding feature Di , i∈1,2,3,4:

[0054]

[0055] Among them, D i represents the decoding features of the i-th stage, represents 1x1 convolution, θ i，D Represents the corresponding parameters. The initial predicted segmentation result P is obtained by upsampling and softmax activation.

[0056] Then, the distance transform technique (DT) is applied to further utilize the structural consistency for prediction correction. Specifically, during the network supervised training process, the prediction result and the corresponding true segmentation template are first transformed as follows by distance transform:

[0057]

[0058] Among them, s l Represents the predicted value or true value of pixel l, l x and l y represents the spatial position of pixel l, k x and k y represents the spatial position of the background pixel k closest to pixel l, dist M (l, k) represents the distance between pixels l and k (using measure M), I fg Represents the set of foreground pixels in the image, I bg Represents the background pixel set in the image. Through the above DT mapping, the regional structural information around each pixel is effectively integrated together, so the KL divergence (relative entropy) constraint is used to predict the image I after DT transformation. D(P) And the DT transformed image I corresponding to the true template D(G) The consistency of distribution is maintained as follows:

[0059] D SCR =D KL (I D(P) ||I D(G) ) (9)

[0060] By applying the above consistency constraints, the structural information around the pixels is effectively utilized to supplement the single pixel-level supervised learning, thereby effectively improving the accuracy of image segmentation.

[0061] It should be noted that, without violating the spirit and scope of the present invention, the technical field of the present invention may make appropriate changes or modifications to the above embodiments. For example, the encoding network includes more encoding stages. For another example, in the decoding stage, it is not necessary to use the structural consistency correction method. In the case where the structural consistency correction method is preferably included, the present invention is also referred to as a medical image segmentation method based on focusing-aggregation-correction context learning. The trained convolutional neural network can be used for real-time medical image segmentation. In addition, the method of the present invention can further integrate the design of semi-supervised and weakly supervised techniques to better handle a large number of unlabeled and weakly labeled samples.

[0062] To further verify the effect of the present invention, experiments were conducted. The results showed that the best segmentation effect was achieved in the experiment of segmenting lesion areas based on lung CT images of patients infected with the new coronavirus. The average intersection over union (mIoU) index of single-class lesion segmentation reached 73.2%, and the average mIoU of multi-class lesion joint segmentation reached 40.0%, which exceeded the existing deep learning model based on this task.

[0063] In summary, the present invention breaks through the deficiency of traditional deep learning methods that only use high-level features of the full convolutional network for decoding, uses autofocus and panoramic aggregation modules for integration, effectively utilizes the contextual information of the same level and cross-level of the encoding network, captures richer fine-grained feature representations, and realizes the complementarity of feature information at different levels; the designed bidirectional gated context aggregation method makes full use of the same-level encoding features in the autofocus module; the designed panoramic aggregation module adaptively combines the cross-level encoding features, effectively captures the long-range dependencies of the context, and realizes the information complementarity between features at different levels; the designed structural consistency correction method, as a supplement to traditional pixel-level supervised learning, improves image segmentation performance through constraints based on structural relationships. In short, the present invention effectively encodes the same-level features through the autofocus module, realizes the efficient fusion of cross-level features through the panoramic aggregation module, and finally further improves the accuracy of image segmentation through the structural consistency correction method.

[0064] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0065] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0066] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0067] The computer program instructions for performing the operation of the present invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed entirely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be personalized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present invention.

[0068] Various aspects of the present invention are described herein with reference to the flow charts and / or block diagrams of the methods, devices (systems) and computer program products according to embodiments of the present invention. It should be understood that each box of the flow chart and / or block diagram and the combination of each box in the flow chart and / or block diagram can be implemented by computer-readable program instructions.

[0069] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processor of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0070] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operating steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0071] The flow charts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the systems, methods and computer program products according to multiple embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a part of a module, a program segment or an instruction, and a part of the module, a program segment or an instruction contains one or more executable instructions for realizing the specified logical function. In some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented by a dedicated hardware-based system that performs the specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions. It is well known to those skilled in the art that it is equivalent to implement it by hardware, implement it by software, and implement it by combining software and hardware.

[0072] Embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the marketplace, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein. The scope of the present invention is defined by the appended claims.

Claims

1. A contextual learning medical image segmentation method based on focus fusion, comprising the following steps: Preprocessing the original image data; The preprocessed image data is input into a trained convolutional neural network for feature extraction, wherein the encoding network of the convolutional neural network comprises a plurality of encoding stages and each encoding stage is sequentially connected to an automatic focusing module and a panoramic aggregation module, and the decoding network of the convolutional neural network sequentially decodes the features of the encoding stages and generates corresponding decoding features, wherein the automatic focusing module is used for multi-scale contextual information mining of the same-level encoding features, and the panoramic aggregation module is used for aggregating cross-level encoding features from different encoding stages; Wherein, the automatic focusing module performs the following steps: The defined multi-level transformation operations are performed on the encoding features of each encoding stage to obtain the corresponding feature map, which is expressed as: in, Represents a series of continuous transformation operations defined, Indicates the corresponding parameter, f i represents the encoding features of encoding stage i, represents f i The feature map obtained after the tth transformation; The automatic focusing module uses bidirectional gated context aggregation in the same-level encoding process to automatically aggregate the features in the multi-level encoding mapping of the same-level stage, and the calculation formula is: in and Represents three gates, which are used to adaptively select the flow information passing through, σ represents the sigmoid activation function, represents pixel-by-pixel multiplication, represents a 1x1 convolutional layer, represents the corresponding convolution parameters, is the context controller, It is a memory; The bidirectional gated context aggregation is a bidirectional aggregation process, which is expressed as: in Represents the forward integration process The output features of Reverse integration process The output features, F i，AM It represents the feature output after bidirectional integration at the same level of the autofocus module; The panoramic aggregation module performs the following steps: Concatenate and aggregate the features of each encoding stage to obtain the global context information Z; The attention mechanism is used to guide the global context information into the encoding features of each encoding stage, so as to supplement the information of each encoding stage with global information, which can be expressed as: in, represents the self-attention mechanism, AvgPool represents the global average pooling operation, σ represents the sigmoid activation function, and F i，AM represents the encoding features after the automatic focusing module at encoding stage i, θ i，Z Indicates the corresponding parameters, Indicates the splicing process; The panoramic aggregation module integrates the encoding features of different encoding stages into the current stage, which is expressed as: r i =1-σ(F i，AM ) Among them, r i Indicates that the gate function is used to selectively control the transmission of information. represents pixel-by-pixel multiplication, represents the cross-level feature aggregation unit, F i,PM represents the output feature map of the panorama aggregation module at the i-th stage; The decoding network generates decoding features based on structural consistency correction and performs the following steps: Generate initial predicted segmentation results; During the network training process, the predicted segmentation result and the corresponding true segmentation template are transformed as follows through distance transformation: Among them, s l Represents the predicted value or true value of pixel l, l x and l y represents the spatial position of pixel l, k x and k y represents the spatial position of the background pixel k closest to pixel l, dist M (l, k) represents the distance between pixels l and k, I fg Represents the set of foreground pixels in the image, I bg Represents the set of background pixels in the image; Use KL divergence constraint to predict the distribution consistency between the distance transformed image of the segmentation result and the distance transformed image of the corresponding true segmentation template; The generating of the initial prediction segmentation result includes: The decoding network sequentially decodes the features of the encoding stage and generates the corresponding decoding features D i ; The initial predicted segmentation result is obtained by upsampling and softmax activation.

2. The method according to claim 1, characterized in that The encoding network sequentially comprises a plurality of residual convolution modules.

3. A computer-readable storage medium having a computer program stored thereon, wherein: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 2 are implemented.

4. A computer device comprising a memory and a processor, wherein a computer program capable of running on the processor is stored in the memory, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 2 are implemented.

Images