Colorectal tumor image segmentation method and system based on automatic receptive field and hierarchical graph interaction

Abstract

The invention provides a colorectal tumor image segmentation method and system based on automatic receptive field and hierarchical graph interaction Transformer, and the method comprises the steps: employing BiFormer as a backbone network, carrying out the extraction of spatial features and long-distance dependence features of a colorectal tumor image, and obtaining four different levels of colorectal tumor coding features; sequentially inputting the four levels of colorectal tumor coding features into an automatic receptive field module; embedding feature representations generated in adjacent automatic receptive field modules into a potential graph structure, and modeling semantic relationship dependence among different regions; the constructed graph structure is input into a hierarchical graph interaction Transform module; and designing a decoder network based on a confidence aggregation feature fusion module. According to the method, the capability of sensing the tumor area content of the model can be enhanced by utilizing the automatic receptive field and hierarchical graph interaction, and the accuracy and robustness of colorectal tumor image segmentation are improved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and system for colorectal tumor image segmentation based on automatic receptive field and hierarchical graph interaction. Background Technology

[0002] With the development of artificial intelligence (AI) technology, medical image analysis technology has made significant progress. By using AI methods to perform refined segmentation of colonoscopy images, tumor tissue structures can be identified quickly and accurately, providing a scientific basis for the early detection and diagnosis of colorectal cancer. Simultaneously, these technologies will provide strong support for doctors in clinical diagnosis and treatment, improving the accuracy of disease monitoring and prognostic guidance. Utilizing technologies such as computer vision and deep learning to conduct automated colorectal tumor image segmentation research can effectively improve the accuracy of doctors' image data analysis, promoting the intelligent and automated progress of smart healthcare.

[0003] In recent years, significant progress has been made in automated polyp segmentation research in colonoscopy images. Early studies mainly analyzed polyp features by examining color distribution, texture characteristics, and structural variations, and extracted hand-designed features accordingly to distinguish polyps from the background. Thanks to the ability of hand-designed features to identify specific patterns, these studies were effective in certain scenarios. However, they performed poorly in challenging scenarios, such as polyps with blurred boundaries, background interference, and variations in size / shape. In recent years, convolutional neural network (CNN)-based methods have shown significant advantages in medical image segmentation, especially in the extraction and integration of target features. Most CNN methods employ an encoder-decoder framework, where the encoding stage extracts high-level semantics and low-level spatial details, and the decoding stage integrates these features to generate a predictive mask. Existing CNN-based segmentation methods are generally divided into general-purpose and specialized categories, depending on their application scope. Research indicates that accurate tumor identification and intelligent guidance during surgery are major challenges affecting tumor diagnosis and treatment.

[0004] Most existing colorectal tumor image segmentation methods primarily focus on size variations, boundary curves, background regions, dense skip connections, and dense residual scale fusion, which can improve performance. However, these adjustments are typically achieved by adding extra layers and explicitly expanding the network and its connections. This increases model complexity, potentially negatively impacting generalization ability from test samples with similar distributions, and requires large datasets. Furthermore, it may affect images where polyp sizes are not representative. Therefore, there is an urgent need for a colorectal tumor image segmentation method that supports the perception of multi-scale features and possesses hierarchical interactive perceptual understanding capabilities. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the shortcomings of the existing technology by providing a method and system for colorectal tumor image segmentation based on automatic receptive field and hierarchical graph interaction, so as to improve the intelligence and accuracy of tumor image segmentation, thereby improving the accuracy of early diagnosis of colorectal cancer, helping doctors to more accurately identify and locate tumors and their precursor lesions, thereby reducing the rate of missed diagnosis and misdiagnosis, improving the survival rate of patients, optimizing clinical treatment plans, and reducing unnecessary surgery and treatment procedures.

[0006] The technical problem to be solved by the present invention is achieved through the following technical solution: The first aspect of the present invention provides a method for colorectal tumor image segmentation based on automatic receptive field and hierarchical graph interaction Transformer, the method comprising the following steps:

[0007] Step S1: Using BiFormer as the backbone network, spatial features and long-distance dependency features are extracted from colorectal tumor images to obtain four different levels of colorectal tumor coding features.

[0008] Step S2: Input the four-level colorectal tumor coding features into the automatic receptive field module in sequence to retain the useful information of each level to the maximum extent and generate a more discriminative representation;

[0009] Step S3: Embed the feature representations generated in adjacent automatic receptive field modules into a latent graph structure to model the semantic relationship dependencies between different regions;

[0010] Step S4: Input the constructed graph structure into the hierarchical graph interaction Transformer module to extract the interaction enhancement features between adjacent levels, thereby enriching the visual semantic information of the features;

[0011] Step S5: Design a decoder network based on the confidence aggregation feature fusion module, fuse the interaction enhancement features extracted by the hierarchical graph interaction module, and refine the image details to obtain the final colorectal tumor image segmentation result.

[0012] Furthermore, the method for extracting colorectal tumor coding features in step S1 is as follows: using BiFormer as the backbone network, spatial features and long-distance dependency features are extracted. Based on the main network of BiFormer, the model is allowed to focus on different parts of the colorectal tumor image through a multi-layer dynamic sparse attention mechanism and a feedforward neural network, so as to achieve more content-aware and more flexible allocation of computing resources, and obtain the coding features of the colorectal tumor image at four levels.

[0013] Furthermore, the automatic receptive field module in step S2 consists of convolution operations and search cells. A search cell can be viewed as a directed acyclic graph composed of multiple nodes connected by directed dashed lines. Each dashed line represents a search operation in the search space. Each node sums the outputs of all its predecessor nodes through these connections.

[0014] Furthermore, the automatic receptive field module is constructed by using a search unit to automatically find suitable feature enhancement operations in each layer, achieving adaptive adaptation to differences between different layers under task-driven conditions, maximizing the retention of useful information at each layer, and generating more discriminative representations for each layer.

[0015] Furthermore, the latent graph structure is constructed in step S3 as follows: First, the feature size of the feature representations generated in the automatic receptive field modules at different levels is unified and reshaped; then, the reshaped features are input into two convolution operations to reduce the dimensionality and perform feature aggregation, thereby constructing a latent graph composed of graph nodes and a predefined number of vertices.

[0016] Furthermore, the construction method of the hierarchical graph interaction Transformer module in step S4 is as follows: (1) the inner product operation is used as a projection operation, which generates a graph representation in the potential interaction space by performing matrix multiplication in the spatial dimension; (2) the Transformer is used to realize local alignment and information transmission between graphs, so as to model the long-distance dependency relationship between graph nodes and can reproject them back to the coordinate space; (3) through bidirectional graph interaction, the potential graph nodes are enhanced at the same time, thereby realizing more expressive visual semantic mining of tumor targets.

[0017] Furthermore, the decoder network based on the confidence aggregation feature fusion module in step S5 includes: a confidence aggregation feature fusion module and a conventional fusion operation consisting of convolution operation, batch normalization operation, and ReLU function, the input of which includes six interactive feature maps generated by three hierarchical graph interaction modules.

[0018] Furthermore, the decoder network based on the confidence-based feature fusion module is constructed as follows: The confidence-based feature fusion module concatenates and fuses adjacent interactive feature maps to obtain coarse-grained predictions. Then, the features are re-weighted to refine local details in blurred regions. Finally, the fusion operation yields refined colorectal tumor image segmentation results. In addition, during the decoding process, weighted binary cross-entropy loss and weighted IoU loss are used to supervise the model's training.

[0019] A second aspect of the present invention provides a colorectal tumor image segmentation system based on an automatic receptive field and hierarchical graph interaction Transformer, the system comprising:

[0020] The colorectal tumor coding feature unit is used to extract spatial features and long-distance dependency features from colorectal tumor images using BiFormer as the backbone network, resulting in four different levels of colorectal tumor coding features.

[0021] A more discriminative representation unit is generated to sequentially input the four-level colorectal tumor coding features into the automatic receptive field module, so as to retain the useful information of each level to the maximum extent.

[0022] The latent graph structure building unit is used to embed the feature representations generated in adjacent automatic receptive field modules into a latent graph structure, which is then used as input to the hierarchical graph interaction module.

[0023] The Hierarchical Graph Interaction Transformer module building unit is used to input the constructed graph structure into the Hierarchical Graph Interaction Transformer module to extract the interaction enhancement features between adjacent levels, thereby enriching the visual semantic information of the features;

[0024] The decoder network based on the confidence aggregation feature fusion module is used to fuse the interactive enhancement features extracted by the hierarchical graph interaction module and refine image details to obtain the final refined colorectal tumor image segmentation result.

[0025] Compared with the prior art, the present invention has the following advantages:

[0026] (1) By introducing an automatic receptive field module, this invention retains the useful information at each level to the maximum extent. Compared with the traditional method of using only a single Transformer network as an encoder, it has a stronger semantic expression capability and can improve the image segmentation and detection accuracy of colorectal tumors and non-tumor areas.

[0027] (2) The present invention adopts the hierarchical graph interaction Transformer module, which can fully explore the interaction enhancement features of adjacent levels to enrich the visual semantic information of the features and improve the coding ability and robustness of the model in cases where the colorectal tumor images are complex in shape, different in size, and have many types.

[0028] (3) The present invention uses a decoder network based on a confidence-aggregated feature fusion module to fuse the interactive enhancement features extracted by the hierarchical graph interaction module and refine the image details to obtain the final refined colorectal tumor image segmentation result.

[0029] (4) The application of the refined segmentation technology of this invention helps to promote the development of smart healthcare, promote the efficient use of medical resources, reduce medical costs, improve public health, promote the application of medical artificial intelligence, and has significant social value in improving public health. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram of the process of colorectal tumor image segmentation using the Automatic Receptive Field and Hierarchical Graph Interaction Transformer provided in an embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram of the automatic receptive field module construction process provided in an embodiment of the present invention;

[0033] Figure 3 This is a schematic diagram of the construction process of the hierarchical graph interaction Transformer module provided in the embodiment of the present invention;

[0034] Figure 4 This is a schematic diagram of the construction process of the confidence aggregation feature fusion module provided in the embodiment of the present invention;

[0035] Figure 5 This is a schematic diagram of the decoder network construction process based on the confidence aggregation feature fusion module provided in the embodiment of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Reference Figure 1 This embodiment provides a method for colorectal tumor image segmentation based on an automatic receptive field and hierarchical graph interaction Transformer, to improve the accuracy of colorectal tumor image segmentation. The method includes the following steps:

[0038] S100. Using BiFormer as the backbone network, spatial features and long-distance dependency features are extracted from colorectal tumor images to obtain four different levels of colorectal tumor coding features.

[0039] As one implementation of step S100, to enhance the model's ability to understand image semantics, a BiFormer model is introduced to extract encoded features from the input image. Specifically:

[0040] Suppose the input colorectal tumor image is H and W represent the height and width of the image, respectively. The image is then linearly mapped and processed through a series of Transformer blocks to generate four different levels of colorectal tumor coding features {F1′, F2′, F3′, F4′}.

[0041] S200. Input the four-level colorectal tumor coding features {F1′,F2′,F3′,F4′} into the automatic receptive field module in sequence to retain the useful information of each level to the maximum extent and generate a more discriminative representation {F1,F2,F3,F4}. Figure 2 A schematic diagram of the automatic receptive field module construction process provided in an embodiment of the present invention is shown.

[0042] As one implementation of step S200, due to the inherent high similarity between the tumor target and the background environment, directly using a self-attention mechanism to establish long-range dependencies inevitably introduces irrelevant information from background interference, resulting in poor tumor target segmentation. Therefore, the automatic receptive field module aims to improve the understanding and discrimination ability of target regions or attributes in the image, and to discover potentially distinguishable regions. Specifically, the automatic receptive field module is constructed as follows:

[0043] First, unlike traditional search units that use a single 1×1 convolution for dimensionality compression before inputting features into a densely connected search structure, the automatic receptive field module adopts an Inception-style design. It uses a multi-branch structure for 1×1 convolutional dimensionality reduction, thus avoiding information loss caused by drastic compression in a single channel. These independent branches form multiple information streams, achieving feature decoupling. Then, the decoupled features are fused to generate a more discriminative feature representation.

[0044] Each branch of the automatic receptive field module consists of a search unit containing two intermediate nodes. The search operation is performed independently within each branch to maximize the retention of useful information. A search unit can be viewed as a directed acyclic graph consisting of multiple nodes. For example... Figure 2As shown in search cell-2, nodes are connected by dashed arrows, with each dashed line representing a search operation S in the search space O. Each node sums the outputs of all its predecessor nodes through these connecting lines. To make the search space continuous, the operation selection is relaxed from a discrete form to the sum of the softmax weights of the candidate operations, denoted as the learnable structural parameter α:

[0045]

[0046] Where op(·) represents an operation in the search space, S (i,j) (·) indicates that the node I is from i To node I j The search operation. After the search process is complete, S (i,j) The (·) operation will be replaced by the operation in O corresponding to the maximum value of α, op. For each node, retain the edges with the highest probability among all input edges, the number of which is equal to the number of input nodes of the search unit, to simplify the network structure.

[0047] S300: Embed the feature representations generated in adjacent automatic receptive field modules into a latent graph structure to model the semantic relationship dependencies between different regions.

[0048] As one implementation of step S300, the latent graph structure construction is based on the output from step S200. First, a more discriminative representation F is generated for the data passed through the automatic receptive field module. i (i = 1, 2, 3, 4), bilinear interpolation is used to unify the feature sizes generated in multiple automatic receptive field modules to H′×W′×C′, reshaping them into... Where C′ represents the unified number of channels, and L=H′×W′ represents the spatial resolution. Then, the reshaped feature F i The input will be two 1×1 convolutional layers, denoted as ′. and This achieves dimensionality compression and cross-channel feature aggregation. Then, the latent graph is constructed as follows:

[0049]

[0050] in, This represents the node features in the graph structure, where N is the preset number of nodes. The inner product operation can be viewed as a projection operation, which maps the feature map to the latent interaction space by performing matrix multiplication along the spatial dimension, thereby generating graph node features. Through this operation, the feature map F in the original coordinate space is obtained. i Embedded into a more compact representation It helps to achieve more consistent identification of regions with similar visual semantics but different locations in the latent space.

[0051] S400. Input the constructed graph structure into the hierarchical graph interaction Transformer module to extract the interaction enhancement features between adjacent levels, thereby enriching the visual semantic information of the features. Figure 3 This diagram illustrates the construction process of the hierarchical graph interaction Transformer module provided in an embodiment of the present invention.

[0052] As one implementation of step S400, the hierarchical graph interaction Transformer module comprises two parts: a hierarchical graph interaction mechanism and Transformer learning and remapping. The hierarchical graph interaction mechanism is based on obtaining the latent graph structure through projection of the latent space graph to achieve local alignment and information transfer between graphs. Specifically, for the graph nodes of the i-th stage... Graph nodes in stage i+1 Two-way interaction is achieved using nonlocal operations with softmax:

[0053]

[0054] in, Indicates from arrive The interaction weight matrix, S (i+1)→i This represents the interaction in the opposite direction. The functions ψ1(·), ψ2(·), θ1(·), and θ2(·) are all learnable graph node feature transformation operations. The above interaction matrix can be viewed as an alignment matrix between nodes, used to measure the correlation between nodes in different levels of the graph, in order to mine complementary visual semantic information between hierarchical feature maps. Simultaneously, the graph nodes... and By splicing the data and compressing the channel dimensions, a joint graph representation is obtained. Then, by comparing this potential graph representation with S respectively i→(i+1) With S (i+1)→i Multiplication enables interactive updates between graphs:

[0055]

[0056] in, This represents the enhanced graph node features through bidirectional interaction. Through this interaction mechanism, potential graph nodes... and It can achieve synergistic enhancement, thereby more effectively mining the visual semantic information of tumor targets.

[0057] As one implementation of step S400, the Transformer learning and remapping aims to model interactive graph nodes. and Higher-order dependencies between them. For example, first, the adjacency matrix A and its corresponding diagonal matrix D are obtained by calculating their inner product. Then, the graph structure information is embedded into the Transformer module as a Laplacian matrix:

[0058]

[0059] in, Represents interaction graph nodes The Laplace matrix. Then, and The standard Transformer module is input together, and feature modeling and enhancement are performed through a multi-head self-attention mechanism:

[0060]

[0061] in, This enhances the graph node features. Finally, the reprojection function is used. This is mapped back to the original coordinate space to restore its spatial structure. Finally, a function with a 1×1 convolution is used. Reprojected visual semantic features Compared with the original feature map F i To perform the fusion, the expression is as follows:

[0062]

[0063] Among them, F (i+1)→i This represents the feature map of layer i after the layer (i+1) is enhanced by the Transformer block of layer i.

[0064] S500: Design a decoder network based on the confidence aggregation feature fusion module, fuse the interaction enhancement features extracted by the hierarchical graph interaction module, and refine the image details to obtain the final colorectal tumor image segmentation result. Figure 4 This is a schematic diagram of the construction process of the confidence aggregation feature fusion module provided in the embodiment of the present invention. Figure 5 This is a schematic diagram of the decoder network construction process based on the confidence aggregation feature fusion module provided in the embodiment of the present invention.

[0065] As one implementation of step S500, the specific process of constructing the confidence-based aggregated feature fusion module is as follows:

[0066] First, such as Figure 4 As shown, feature map F(i+1)→i and F i→(i+1) The features are then concatenated and compressed to obtain the features under a single-stage interaction. Then, the spliced ​​features A coarse-grained prediction map P is obtained by fusing the results through a series of convolutional modules. i :

[0067]

[0068] Wherein, Conv(·) represents a layer consisting of a convolutional layer, batch normalization (BN), and ReLU.

[0069] Sigmoid(·) represents the Sigmoid activation function. However, this coarse-grained prediction may lead to undesirable segmentation results in blurry regions because the pixel-by-pixel prediction scores in these regions often lack confidence. Therefore, by using P... i With 1-P i Multiplication explicitly emphasizes blurred areas and reweights the stitched features for refinement, as shown in the following formula:

[0070]

[0071] In this context, the symbol ⊙ represents element-wise multiplication. P i ⊙(1-P i ) is considered a pixel-by-pixel blur indicator: if P i If P is close to 0.5, then P i ⊙(1-P i The product value is larger; conversely, if P... i If P is close to 0 or 1, then i ⊙(1-P i The product value approaches zero. This is achieved by comparing this term with the feature F. i C Multiplication and the introduction of skip connections enhance feature aggregation in blurred boundary regions while preserving information from high-confidence regions. Then, a fusion module (consisting of a Conv-BN-ReLU layer and a sigmoid function) is used to obtain the final refined prediction result.

[0072]

[0073] Among them, such as Figure 5 As shown, P i ′ represents the refined prediction output of the i-th confidence aggregation feature fusion module, corresponding to the intermediate features. The expression is as follows:

[0074]

[0075] Where Conv(·) represents the feature concatenation operation. The feature map after reweighting. It is an intermediate feature output by the previous level fusion module.

[0076] Based on the above operations, the refined prediction output of the third confidence-based feature fusion module is used as the final refined segmentation result of colorectal tumors.

[0077] Another embodiment of the present invention provides a colorectal tumor image segmentation system based on an automatic receptive field and hierarchical graph interaction Transformer, the system being used to perform the method steps described in the foregoing method embodiments. The system includes:

[0078] The colorectal tumor coding feature unit is used to extract spatial features and long-distance dependency features from colorectal tumor images using BiFormer as the backbone network, resulting in four different levels of colorectal tumor coding features.

[0079] A more discriminative representation unit is generated to sequentially input the four-level colorectal tumor coding features into the automatic receptive field module, so as to retain the useful information of each level to the maximum extent.

[0080] The latent graph structure building unit is used to embed the feature representations generated in adjacent automatic receptive field modules into a latent graph structure, which is then used as input to the hierarchical graph interaction module.

[0081] The Hierarchical Graph Interaction Transformer module building unit is used to input the constructed graph structure into the Hierarchical Graph Interaction Transformer module to extract the interaction enhancement features between adjacent levels, thereby enriching the visual semantic information of the features;

[0082] The decoder network based on the confidence aggregation feature fusion module is used to fuse the interactive enhancement features extracted by the hierarchical graph interaction module and refine image details to obtain the final refined colorectal tumor image segmentation result.

[0083] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0084] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for colorectal tumor image segmentation based on automatic receptive field and hierarchical graph interaction, characterized in that: Includes the following steps, Step S1: Using BiFormer as the backbone network, spatial features and long-distance dependency features are extracted from colorectal tumor images to obtain four different levels of colorectal tumor coding features. Step S2: Input the four-level colorectal tumor coding features into the automatic receptive field module in sequence to retain the useful information of each level to the maximum extent and generate a more discriminative representation; Step S3: Embed the feature representations generated in adjacent automatic receptive field modules into a latent graph structure to model the semantic relationship dependencies between different regions; Step S4: Input the constructed graph structure into the hierarchical graph interaction Transformer module to extract the interaction enhancement features between adjacent levels, thereby enriching the visual semantic information of the features; Step S5: Design a decoder network based on the confidence aggregation feature fusion module, fuse the interaction enhancement features extracted by the hierarchical graph interaction module, and refine the image details to obtain the final colorectal tumor image segmentation result.

2. The method for colorectal tumor image segmentation based on automatic receptive field and hierarchical graph interaction according to claim 1, characterized in that: The extraction method of colorectal tumor coding features in step S1 is as follows: using BiFormer as the backbone network, spatial features and long-distance dependency features are extracted. Based on the main network of BiFormer, the model is allowed to focus on different parts of the colorectal tumor image through a multi-layer dynamic sparse attention mechanism and a feedforward neural network, so as to achieve more content-aware and more flexible allocation of computing resources, and obtain the coding features of colorectal tumor images at four levels.

3. The method for colorectal tumor image segmentation based on automatic receptive field and hierarchical graph interaction according to claim 1, characterized in that: The automatic receptive field module in step S2 consists of convolution operations and search cells. A search cell can be viewed as a directed acyclic graph composed of multiple nodes connected by directed dashed lines. Each dashed line represents a search operation in the search space. Each node sums the outputs of all its predecessor nodes through these connections.

4. The method for colorectal tumor image segmentation based on automatic receptive field and hierarchical graph interaction according to claim 3, characterized in that: The automatic receptive field module is constructed by using a search unit to automatically find suitable feature enhancement operations in each layer, achieving adaptive adaptation to differences between different layers under task-driven conditions, preserving useful information at each layer to the maximum extent, and generating more discriminative representations for each layer.

5. The method for colorectal tumor image segmentation based on automatic receptive field and hierarchical graph interaction according to claim 1, characterized in that: The latent graph structure in step S3 is constructed as follows: First, the feature size of the feature representations generated in the automatic receptive field modules at different levels is unified and reshaped; then, the reshaped features are input into two convolution operations to reduce the dimensionality and perform feature aggregation, thereby constructing a latent graph composed of graph nodes and a predefined number of vertices.

6. The method for colorectal tumor image segmentation based on automatic receptive field and hierarchical graph interaction according to claim 1, characterized in that: The construction method of the hierarchical graph interaction Transformer module in step S4 is as follows: (1) The inner product operation is used as a projection operation, which generates a graph representation in the potential interaction space by performing matrix multiplication in the spatial dimension; (2) Use Transformer to achieve local alignment and information transfer between graphs, so as to model long-distance dependencies between graph nodes and reproject them back to coordinate space; (3) Through bidirectional graph interaction, potential graph nodes are enhanced simultaneously, thereby enabling more expressive visual semantic mining of tumor targets.

7. The method for colorectal tumor image segmentation based on automatic receptive field and hierarchical graph interaction according to claim 1, characterized in that: The decoder network based on the confidence aggregation feature fusion module in step S5 includes: the confidence aggregation feature fusion module and a conventional fusion operation consisting of convolution operation, batch normalization operation and ReLU function. Its input includes six interactive feature maps generated by three hierarchical graph interaction modules.

8. The method for colorectal tumor image segmentation based on automatic receptive field and hierarchical graph interaction according to claim 7, characterized in that: The decoder network based on the confidence-based feature fusion module is constructed as follows: The confidence-based feature fusion module concatenates and fuses adjacent interactive feature maps to obtain coarse-grained predictions. Then, the features are re-weighted to refine local details in blurred regions. Finally, the fusion operation yields refined colorectal tumor image segmentation results. Furthermore, during the decoding process, weighted binary cross-entropy loss and weighted IoU loss are used for model training supervision.

9. A colorectal tumor image segmentation system based on automatic receptive field and hierarchical graph interaction, characterized in that: The system includes, The colorectal tumor coding feature unit is used to extract spatial features and long-distance dependency features from colorectal tumor images using BiFormer as the backbone network, resulting in four different levels of colorectal tumor coding features. A more discriminative representation unit is generated to sequentially input the four-level colorectal tumor coding features into the automatic receptive field module, so as to retain the useful information of each level to the maximum extent. The latent graph structure building unit is used to embed the feature representations generated in adjacent automatic receptive field modules into a latent graph structure, which is then used as input to the hierarchical graph interaction module. The Hierarchical Graph Interaction Transformer module building unit is used to input the constructed graph structure into the Hierarchical Graph Interaction Transformer module to extract the interaction enhancement features between adjacent levels, thereby enriching the visual semantic information of the features; The decoder network based on the confidence aggregation feature fusion module is used to fuse the interactive enhancement features extracted by the hierarchical graph interaction module and refine image details to obtain the final refined colorectal tumor image segmentation result.

Images