Polyp image segmentation method, device and medium based on layer interaction and diffusion refinement

Abstract

The application belongs to the technical field of deep learning and image processing, and relates to a polyp image segmentation method, device and medium based on layer interaction and diffusion refinement. The method first extracts multi-scale hierarchical features of a polyp image by using a backbone network, obtains detailed features by a detailed feature extraction module, and obtains a coarse mask by a hierarchical interaction fusion module. Then, a diffusion U-shaped network is trained by performing probabilistic transition feature noise addition training based on the difference between the coarse mask and a real label. The noise-added mask and the detailed features are input into a parallel residual coding module and the diffusion U-shaped network for training. Finally, in the inference stage, the coarse mask and the detailed features are taken as conditions, the diffusion U-shaped network is used for denoising refinement, and local pixel refinement is performed on a low-fine probability area to obtain a final fine segmentation result. The application can enhance the multi-level feature interaction capability, improve the refinement capability of unreliable pixel areas of the coarse mask, and thus improve the colon polyp image segmentation precision.

Description

Technical Field

[0001] This invention belongs to the fields of deep learning and image processing technology, and relates to a method, apparatus and medium for polyp image segmentation based on layer interaction and diffusion refinement. Background Technology

[0002] Polyp image segmentation is a key task in intelligent analysis of endoscopic medical images. The segmentation results can be used for polyp region localization, boundary extraction, and subsequent auxiliary diagnosis. In recent years, deep learning technology has been widely used in image segmentation, especially in polyp image segmentation. Common methods typically extract multi-level features through a backbone network and then combine encoding, decoding, or multi-scale fusion strategies to complete segmentation prediction.

[0003] However, existing polyp image segmentation methods still have shortcomings when dealing with complex backgrounds, blurred boundaries, and polyp targets with diverse shapes. On the one hand, inconsistencies in semantic representation at different levels are prone to occur during multi-scale feature fusion, making it difficult for high-resolution features to simultaneously preserve semantic information and boundary details. On the other hand, unreliable pixel regions are often present in the initial coarse segmentation results. Without a progressive correction mechanism for the differences between the coarse segmentation results and the true labels, the boundary integrity and precision of the final segmentation results remain limited. Therefore, a polyp image segmentation method that can enhance the interaction capabilities of hierarchical features and perform progressive refinement on the coarse segmentation results is needed. Summary of the Invention

[0004] To address the aforementioned technical problems in the existing technology, this invention proposes a polyp image segmentation method, apparatus, and medium based on layer interaction and diffusion refinement, the specific technical solution of which is as follows:

[0005] A polyp image segmentation method based on layer interaction and diffusion refinement is proposed. The method employs a polyp image segmentation model, which includes: a backbone network for extracting multi-scale hierarchical features from the input polyp image; a detail feature extraction module, comprising feature branches and residual bypasses, which take input feature T1 and sequentially perform 1×1 convolution, Sigmoid gate weight calculation, element-wise multiplication with T1, and 3×3 convolution in the feature branches. The output of the 3×3 convolution is then added to the output of the residual bypass to obtain the detail feature map; and a layer interaction fusion module performs inter-layer interaction fusion of multi-level features T2, T3, and T4 to obtain a coarse segmentation result. The system consists of a coarse mask and a parallel residual coding module. The coarse mask is jointly encoded with the detail feature map, enabling feature transformation at different scales, inversion enhancement, and residual fusion to obtain residual coded features. A diffuse U-shaped network takes the residual coded features as conditional input and uses a refinement mechanism based on pixel fine probability and transition probability to denoise the coarse mask pixel by pixel, outputting a refined segmentation result. When training the polyp image segmentation model, a probabilistic transition noise addition process is used to enable the diffuse U-shaped network to learn the difference between the coarse mask and the true label. During inference, local refinement is only performed again on regions with low pixel fine probability to obtain the final polyp segmentation result.

[0006] Furthermore, the backbone network employs a pyramid visual transformer (PVT).

[0007] Furthermore, the detail feature extraction module performs information filtering on the lowest-level feature T1. The output of the 1×1 convolution is activated by Sigmoid to obtain the gate weights. The gate weights are multiplied element-wise by T1. The result is then convolved by 3×3 and added to the residual of T1 to obtain the detail feature map.

[0008] Furthermore, the hierarchical interactive fusion module is a three-level residual fusion architecture composed of a pooling residual enhancement unit and a progressive feature convolution fusion. The pooling residual enhancement unit first performs 1×1 convolution processing on features at each scale except for low-level features, then extracts edge features through average pooling at two different scales, and concatenates the obtained edge features to obtain aggregated features. The aggregated features are then further enhanced by 1×1 convolution and feature inversion enhancement and residual connection to obtain further enhanced features. The progressive feature convolution fusion operation is to perform layer-by-layer convolution and upsampling concatenation fusion operations on the enhanced features to obtain the coarse mask.

[0009] Furthermore, the parallel residual coding module first concatenates the detail feature map and the coarse mask element by element to obtain feature F, and then performs feature transformation at different scales through KAN adaptive pooling blocks to obtain pooling features at different scales. The adaptive average pooling output size is set to 1, 3 and 5.

[0010] Then, the Sigmoid activation function is used to invert and enhance the pooling features at each scale, and the enhanced features at different scales are obtained by residual connection with the pooling features at the corresponding scale.

[0011] Finally, the enhanced features at different scales are fused using a stepwise residual fusion method to obtain residual coding features.

[0012] Furthermore, the diffused U-shaped network first obtains the initial denoising prediction result based on the input residual coding features. ;

[0013] Then, the fine probability is calculated based on the initial denoising prediction results. and transition probability And based on fine probability and transition probability Iteratively update the fine probability of the current pixel Based on the current pixel fine probability Perform pixel segmentation decisions to obtain block regions. And sample and update the segmentation results The mathematical expressions for each parameter are as follows:

[0014] ;

[0015] , Scheduling parameters for controlling the intensity of diffusion refinement;

[0016] = ;

[0017] , It is a randomly sampled variable;

[0018] ;

[0019] Finally, the output result is divided into blocks based on the final pixel fineness probability. Only the blocks with low pixel fineness probability are retained for further local fineness re-refinement. All local fineness re-refinement areas are then integrated to obtain the final fineness result.

[0020] Furthermore, the probabilistic transition noise addition process during training of the polyp image segmentation model is a process of generating a random noise mask based on the coarse mask M and the true label GT, specifically including: calculating the cumulative transition probability based on time step t. The cumulative transition probability From noise scheduling parameters The expression is obtained through gradual changes and is as follows: ;

[0021] Based on cumulative transition probability Generate random binary mask The specific expression is: ;

[0022] By conditionally sampling, the true label GT is interleaved with the coarse segmentation result M to generate a random noise mask that transitions from the true label GT to the coarse mask M. The specific expression is: .

[0023] Furthermore, the training process of the polyp image segmentation model employs weighted texture loss, weighted binary cross-entropy loss, and weighted intersection-union ratio loss for supervision, with the output results of the diffusion U-shaped network being monitored. The loss is composed of weighted binary cross-entropy loss. and texture loss The loss of the coarse mask M is composed of a weighted binary cross-entropy loss. Weighted intersection and comparison loss The total loss function is composed of two parts of the loss, and its expression is: .

[0024] A polyp image segmentation device based on layer interaction and diffusion refinement includes one or more processors for implementing the polyp image segmentation method based on layer interaction and diffusion refinement.

[0025] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the polyp image segmentation method based on layer interaction and diffusion refinement.

[0026] The beneficial effects of this invention are as follows: by enhancing the information interaction between multi-level semantic features through the hierarchical interaction fusion module, the feature representation capability is improved; by probabilistic transition feature noise addition process, the diffusion network can perform targeted learning on the difference between the coarse segmentation result and the real label; by the stepwise denoising and refinement mechanism based on fine probability and transition probability, the ability to correct unreliable pixel regions is improved, thereby improving the accuracy and boundary integrity of polyp image segmentation results. Attached Figure Description

[0027] Figure 1 This is an overall structural diagram of a polyp image segmentation method based on layer interaction and diffusion refinement according to the present invention.

[0028] Figure 2 This is a structural diagram of the hierarchical interactive fusion module of the present invention.

[0029] Figure 3 This is a structural diagram of the pooling residual enhancement module of the present invention.

[0030] Figure 4 This is a structural diagram of the parallel residual coding module of the present invention.

[0031] Figure 5 This is a structural diagram of the detailed feature extraction module in an embodiment of the present invention.

[0032] Figure 6 These are experimental visualization comparison diagrams of embodiments of the present invention.

[0033] Figure 7 A schematic diagram of a polyp image segmentation device based on layer interaction and diffusion refinement in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and technical effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0035] like Figure 1 As shown, a polyp image segmentation method based on layer interaction and diffusion refinement includes the following steps:

[0036] Step 1: Obtain the image of the polyp to be segmented.

[0037] Step 2: Construct a polyp image segmentation model to sequentially perform multi-scale hierarchical feature extraction, low-level feature information filtering, coarse segmentation, masking and noise addition, and fine segmentation on the input polyp image. Specifically, this includes:

[0038] Step 2.1: Input the polyp image into the backbone network, which is a Pyramid Visual Transformer (PVT). Multi-scale hierarchical features are extracted using the PVT. In this embodiment, the backbone network outputs four hierarchical features: T1, T2, T3, and T4.

[0039] Step 2.2: As Figure 5 As shown, the lowest-level feature T1 is input into the detail feature extraction module, which includes a feature branch and a residual bypass. In the feature branch, a 1×1 convolution is first performed on T1 to obtain gated features, and then the gate weights are calculated using the Sigmoid activation function. The gate weights are multiplied element-wise by T1 to obtain the gated detail response. A 3×3 convolution is performed on the gated detail response, and the output of the 3×3 convolution is added to the residual of T1 retained by the residual bypass to obtain the detail feature map G. The detail feature map G is used as the conditional input for subsequent parallel residual coding and diffusion refinement.

[0040] Step 2.3: As Figure 2 and Figure 3As shown, multi-level features are input into the hierarchical interactive fusion module. Through a three-level residual fusion architecture consisting of pooling residual enhancement units and progressive feature convolution fusion, multi-level features are subjected to hierarchical interactive fusion to reduce semantic conflicts in multi-scale feature fusion and obtain a coarse segmentation result M, i.e., a coarse segmentation mask.

[0041] Specifically, the pooling residual enhancement unit first processes the input features. Perform a 1×1 convolution, then extract edge features using average pooling at scales of 3 and 5 respectively. Concatenate the resulting features to obtain aggregated features. The corresponding mathematical expression can be written as:

[0042] ,

[0043] ,

[0044] ;

[0045] In the above formula: It is a 1×1 convolution; This is average pooling with a scale of 3; This is average pooling with a scale of 5; , The edge feature information corresponding to average pooling of 3 and 5; This is an aggregation feature.

[0046] Aggregation features Further enhanced features are obtained by performing 1×1 convolution, feature inversion enhancement, and residual connection. The corresponding mathematical expression is:

[0047] ,

[0048] ,

[0049] ;

[0050] In the above formula: These are features generated through a 1×1 convolution; This indicates feature inversion enhancement. This corresponds to the feature after inversion enhancement.

[0051] Then, the final coarse segmentation result M is obtained through progressive feature convolution fusion, and the corresponding mathematical expression is:

[0052] ,

[0053] ,

[0054] ;

[0055] In the above formula: Indicates upsampling; Use the Sigmoid activation function; It is a 3×3 convolution; and These represent the encoded features and the features after the upsampling operation, respectively, with the subscript numbers indicating the network level.

[0056] Step 2.4: Input the detail feature map G and the coarse segmentation result M into the parallel residual coding module, and perform joint coding on the input through multi-level KAN adaptive inversion enhancement and residual fusion to obtain residual coding features.

[0057] Specifically, such as Figure 4 As shown, the detailed feature map G and the coarse segmentation result M are first concatenated to obtain feature F. Then, KAN adaptive pooling blocks are used to perform feature transformations at different scales to obtain pooling features at different scales. The adaptive average pooling output sizes were set to 1, 3, and 5. Then, the Sigmoid activation function was used to invert and enhance the pooling features at each scale, and residual connections were made with the corresponding scale pooling features to obtain enhanced features at different scales. Subsequently, the enhanced features were fused using a stepwise residual fusion method to obtain residual encoded features. This provides conditional encoding that integrates edge details and coarse segmentation context for the diffuse U-shaped network, enhancing the representation of unreliable pixel regions and subsequent refinement capabilities, thereby improving segmentation accuracy and boundary integrity.

[0058] Step 2.5: Encode residual features Input a diffuse U-shaped network to obtain refined segmentation outputs for training and refinement.

[0059] Step 3: Train the polyp image segmentation model, specifically including:

[0060] Step 3.1: Use a probabilistic transition noise addition process to generate a random noise mask based on the coarse segmentation result M and the true label GT.

[0061] Specifically, the transition probabilities are first defined and a random binary mask is generated, including:

[0062] Calculate the cumulative transition probability based on time step t. The cumulative transition probability From noise scheduling parameters It is obtained through gradual changes, specifically as follows: ;

[0063] Based on cumulative transition probability Generate random binary mask Specifically , ;

[0064] By conditionally sampling, the true label GT is transitionally mixed with the coarse segmentation result M to generate a random noise mask that transitions from the true label GT to the coarse segmentation result M. Specifically .

[0065] Step 3.2: Mask the random noise The detailed feature map G is input into the parallel residual coding module and the diffuse U-shaped network for fine-grained segmentation training.

[0066] In one implementation, the model is supervised during training using weighted texture loss, weighted binary cross-entropy loss, and weighted intersection-union ratio loss. The diffusion U-shaped network outputs the results... The loss is composed of weighted binary cross-entropy loss. and texture loss The loss of the coarse segmentation result M is composed of weighted binary cross-entropy loss. Weighted intersection and comparison loss The total loss function is composed of two parts of the loss, and its expression is: .

[0067] By assigning higher weights to hard pixels, the model's ability to focus on difficult pixel regions can be enhanced.

[0068] Step 3.3: Use the trained diffusion U-shaped network to perform pixel-by-pixel denoising and refinement on the coarse segmentation result M to obtain the refined segmentation result.

[0069] Specifically, the coarse segmentation result M and the detail feature map G are first input into the parallel residual encoding module. and diffusion U-shaped network To obtain the initial denoising prediction results , ,in .

[0070] Then, the fine probability is calculated based on the initial denoising prediction results. and transition probability ,in , , The scheduling parameters for controlling the diffusion refinement intensity are used to control the denoising intensity. This is based on the current refinement probability. and transition probability Update current pixel fine probability , = Based on the current pixel fine probability Perform pixel segmentation decisions to obtain block regions. , , It is a random sampling variable; and the segmentation results are updated by sampling. , .

[0071] Finally, the output result is divided into blocks based on the final pixel fineness probability. Only the blocks with lower pixel fineness probability are retained for further local fineness re-refinement. All local fineness re-refinement areas are then integrated to obtain the final fineness result.

[0072] Step 4: Using the trained polyp image segmentation model, input the polyp image to be segmented into the model and output the polyp segmentation result.

[0073] To demonstrate the effectiveness of this method, a subset of polyp images from classic public polyp segmentation datasets such as CVC-300, Kvasir, ColonDB, ETIS, and ClinicDB were selected to train the complete network. The network was then tested on five datasets: CVC-300, Kvasir, ColonDB, ETIS, and ClinicDB, and the results were compared with mainstream polyp segmentation algorithms in the prior art. The proposed model framework was implemented in PyTorch and trained using an NVIDIA GeForce RTX 3060 SUPER GPU with 8GB of memory. All input images were uniformly set to 352×352, and the Adam optimization algorithm was used to optimize the overall parameters with a learning rate of 0.25e-4. The polyp segmentation test results of this invention and eight other methods on the CVC-ClinicDB and Kvasir datasets are shown in Tables 1 and 2.

[0074] The evaluation metrics include: mDice (average Dice coefficient); mIoU (average crossover / union ratio); and S. α For structural measurement; wF β For weighted F-measure; maxE Φ This is the maximum alignment metric.

[0075] Eight methods: Unet is the classic benchmark network for medical image segmentation; UNet++ is a deep improvement based on UNet; SFA is a task model that focuses on the fusion of spatial and frequency domains; PraNet is a parallel inverse attention network, a classic fully supervised polyp segmentation network model; ACSNet is an anatomical consistency network; SimpleUNet2D is a lightweight reconstruction model; and DiffUnet is a cutting-edge architecture that incorporates a diffusion model.

[0076] Table 1:

[0077]

[0078] Table 2:

[0079]

[0080] The results of polyp segmentation tests of this invention and eight other methods on the CVC-300 and CVC-ColonDB datasets are shown in Tables 3 and 4.

[0081] Table 3:

[0082]

[0083] Table 4:

[0084]

[0085] The test results of this invention and eight other methods on polyp segmentation in the ETIS dataset are shown in Table 5 below.

[0086] Table 5:

[0087]

[0088] Among the five evaluation metrics, mDice and mIoU are commonly used in semantic segmentation tasks, and the closer the value is to 1, the better the segmentation result.

[0089] Figure 6 The diagram shows a visual comparison of the experimental results of the method in this invention and other methods. The results show that the polyp segmentation method proposed in this invention can obtain segmentation results with more accurate boundaries and more complete semantic structure.

[0090] Corresponding to the aforementioned embodiments of the polyp image segmentation method based on layer interaction and diffusion refinement, the present invention also provides embodiments of a polyp image segmentation apparatus based on layer interaction and diffusion refinement.

[0091] See Figure 7 The present invention provides a polyp image segmentation device based on layer interaction and diffusion refinement, comprising one or more processors for implementing the polyp image segmentation method based on layer interaction and diffusion refinement in the above embodiments.

[0092] The embodiments of the polyp image segmentation device based on layer interaction and diffusion refinement of the present invention can be applied to any device with data processing capabilities, such as a computer. The device embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 7 The diagram shown is a hardware structure diagram of any device with data processing capabilities, including the polyp image segmentation device based on layer interaction and diffusion refinement according to the present invention. (Except for...) Figure 7 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0093] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0094] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0095] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the polyp image segmentation method based on layer interaction and diffusion refinement described in the above embodiments.

[0096] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0097] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A polyp image segmentation method based on layer interaction and diffusion refinement, employing a polyp image segmentation model, characterized in that, The model includes: The backbone network extracts multi-scale hierarchical features from the input polyp image, obtaining features at scale levels from low to high: T1, T2, T3, and T4; The detail feature extraction module includes a feature branch and a residual bypass. The feature T1 is input, and a 1×1 convolution, Sigmoid gate weight calculation, element-wise multiplication with T1 and a 3×3 convolution are performed in the feature branch in sequence. The output of the 3×3 convolution is added to the output of the residual bypass to obtain the detail feature map. The hierarchical interactive fusion module performs inter-level interactive fusion of multi-level features T2, T3 and T4 to obtain a coarse mask; The parallel residual coding module jointly encodes the detail feature map and the coarse mask to achieve feature transformation at different scales, inversion enhancement and residual fusion, and obtain residual coded features. The diffuse U-shaped network takes the residual coding features as conditional input and uses a refinement mechanism based on pixel fine probability and transition probability to denoise the coarse mask pixel by pixel, and outputs the refined segmentation result. When training the polyp image segmentation model, a probabilistic transition noise addition process is used to enable the diffuse U-shaped network to learn the difference between the coarse mask and the real label; during inference, local refinement is only performed on regions with low pixel fineness probability to obtain the final polyp segmentation result.

2. The method as described in claim 1, characterized in that, The backbone network is the Pyramid Visual Transformer (PVT).

3. The method as described in claim 1, characterized in that, The detail feature extraction module filters information from the lowest-level feature T1. The output of the 1×1 convolution is activated by Sigmoid to obtain the gate weights. The gate weights are multiplied element-wise by T1. The result is then convolved by 3×3 and added to the residual of T1 to obtain the detail feature map.

4. The method as described in claim 1, characterized in that, The hierarchical interactive fusion module is a three-level residual fusion architecture composed of pooling residual enhancement units and progressive feature convolution fusion. The pooling residual enhancement unit first performs 1×1 convolution processing on features at each scale except for low-level features, then extracts edge features through average pooling at two different scales, and concatenates the obtained edge features to obtain aggregated features. The aggregated features are then enhanced by 1×1 convolution and feature inversion, and connected to the residuals to obtain further enhanced features; The progressive feature convolution fusion operation involves performing layer-by-layer convolution and upsampling concatenation and fusion operations on the enhanced features to obtain the coarse mask.

5. The method as described in claim 1, characterized in that, The parallel residual coding module first concatenates the detail feature map and the coarse mask to obtain feature F, and then performs feature transformation at different scales through KAN adaptive pooling blocks to obtain pooling features at different scales. The adaptive average pooling output size is set to 1, 3 and 5. Then, the Sigmoid activation function is used to invert and enhance the pooling features at each scale, and the enhanced features at different scales are obtained by residual connection with the pooling features at the corresponding scale. Finally, the enhanced features at different scales are fused using a stepwise residual fusion method to obtain residual coding features.

6. The method as described in claim 1, characterized in that, The diffused U-shaped network first obtains the initial denoising prediction result based on the input residual coding features. ; Then, the fine probability is calculated based on the initial denoising prediction results. and transition probability And based on fine probability and transition probability Iteratively update the fine probability of the current pixel Based on the current pixel fine probability Perform pixel segmentation decisions to obtain block regions. And sample and update the segmentation results The mathematical expressions for each parameter are as follows: ； , Scheduling parameters for controlling the intensity of diffusion refinement; = ； , It is a randomly sampled variable; ； Finally, the output result is divided into blocks based on the final pixel fineness probability. Only the blocks with low pixel fineness probability are retained for further local fineness re-refinement. All local fineness re-refinement areas are then integrated to obtain the final fineness result.

7. The method as described in claim 1, characterized in that, The probabilistic transition noise addition process during training of the polyp image segmentation model is a process of generating a random noise mask based on the coarse mask M and the ground truth label GT. Specifically, it includes: calculating the cumulative transition probability based on time step t. The cumulative transition probability From noise scheduling parameters The expression is obtained through gradual changes and is as follows: ; Based on cumulative transition probability Generate random binary mask The specific expression is: ; By conditionally sampling, the true label GT is interleaved with the coarse segmentation result M to generate a random noise mask that transitions from the true label GT to the coarse mask M. The specific expression is: .

8. The method as described in claim 1, characterized in that, The training process of the polyp image segmentation model employs weighted texture loss, weighted binary cross-entropy loss, and weighted intersection-over-union (IoU) loss for supervision. The output results of the diffuse U-shaped network are then monitored. The loss is composed of weighted binary cross-entropy loss. and texture loss The loss of the coarse mask M is composed of a weighted binary cross-entropy loss. Weighted intersection and comparison loss The total loss function is composed of two parts of the loss, and its expression is: 。 9. A polyp image segmentation device based on layer interaction and diffusion refinement, characterized in that, It includes one or more processors for implementing the polyp image segmentation method based on layer interaction and diffusion refinement as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the polyp image segmentation method based on layer interaction and diffusion refinement as described in any one of claims 1 to 8.

Images