Colon polyp segmentation method and system based on prototype memory and feature enhancement

Abstract

The invention discloses a colon polyp segmentation method and system based on prototype memory and feature enhancement, and belongs to the field of computer-aided medical diagnos.The method comprises the steps that an edge feature map of a colonoscope image is input into an image segmentation model to be processed, and a segmentation result is obtained; the image segmentation model comprises an input layer, a multi-branch feature enhancement fusion module, a segmentation head and a boundary head which are connected in sequence, the segmentation head and the boundary head are parallel, and the multi-branch feature enhancement fusion module comprises a plurality of branches. Each branch comprises a backbone network, a receptive field module, a discrete wavelet conversion module, a semantic prototype memory bank module and a decoder module which are connected in sequence, and is constructed as follows: transmitting an edge feature map into a multi-branch feature enhancement fusion module to obtain a final fusion feature; and transmitting the final fusion feature into the segmentation head and the boundary head to obtain a segmentation result. A semantic prototype memory bank module and uncertainty estimation are innovatively constructed, the feature storage and utilization capability is improved, and boundary detection is optimized.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided medical diagnosis, and in particular relates to a method and system for segmenting colon polyps based on prototype memory and feature enhancement. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Early screening and accurate treatment of colorectal cancer are crucial, and this highly depends on the proper segmentation of intestinal polyps. According to WHO data, 90% of colorectal cancer patients develop from adenomatous polyps. The size of the intestinal polyp also relates to the risk of malignancy; adenomatous polyps larger than 1 cm have a 37%-55% probability of becoming cancerous within five years. Clinically, the detection of polyps smaller than 5 mm is still insufficient, with a false negative rate exceeding 18%, necessitating improvements in the detection accuracy of polyps with ambiguous boundaries and micropolyps.

[0004] Medical image segmentation networks represent an innovative achievement resulting from the deep integration of artificial intelligence technology and medical practice, possessing significant theoretical and practical value. This technology not only drives the optimization and upgrading of diagnostic and treatment processes but also provides a scientific path for the balanced allocation of medical resources. Traditional medical image analysis, relying on manual interpretation, is inefficient and lacks standardized criteria. However, image segmentation driven by deep learning can automatically process medical images, transforming complex medical images into quantifiable features and indicators. This technological innovation significantly enhances the lesion identification capabilities of medical institutions, offering unique advantages, particularly in the detection of early, minute lesions, thus facilitating early disease diagnosis.

[0005] Currently, there are many polyp segmentation systems based on deep learning architectures in clinical practice (typical examples include U-Net, PraNet, etc.), but they still face difficulties in complex situations, namely, the blurred edges of polyp tissues and the significant heterogeneity in size distribution. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a method and system for segmenting colon polyps based on prototype memory and feature enhancement. It innovatively constructs a semantic prototype memory module and uncertainty estimation, improves feature storage and utilization capabilities, and optimizes boundary detection.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, the present invention provides a method for segmenting colonic polyps based on prototype memory and feature enhancement, comprising: The colonoscopy image to be segmented is acquired and preprocessed to obtain the edge feature map; The edge feature map is input into the image segmentation model for processing to obtain the segmentation result; The image segmentation model comprises an input layer, a multi-branch feature enhancement and fusion module, and parallel segmentation heads and boundary heads, all connected in sequence. The multi-branch feature enhancement and fusion module includes multiple branches, each of which comprises a backbone network, a receptive field module, a discrete wavelet transform module, a semantic prototype memory module, and a decoder module, all connected in sequence. It is constructed as follows: The edge feature map is fed into the multi-branch feature enhancement and fusion module to obtain the final fused feature; The final fused features are fed into the segmentation head and the boundary head to obtain the segmentation result.

[0008] A further technical solution involves preprocessing colonoscopy images using an edge detection algorithm to obtain edge feature maps.

[0009] In a further technical solution, the multi-branch feature enhancement and fusion module is constructed as follows: The edge feature map is fed into the multi-branch backbone network for feature extraction to obtain multi-scale features, which are then fed into the receptive field module. After preprocessing the multi-scale features in the receptive field module, the features are fed into the discrete wavelet transform module for further processing to obtain wavelet enhancement features and boundary features. The wavelet enhancement features are fed into the semantic prototype memory module, combined with the boundary features of the next level scale for enhancement, to obtain multi-scale enhancement features, which are then fed into the decoder. The decoder fuses the multi-scale enhanced features to obtain the final fused features.

[0010] A further technical solution yields wavelet enhancement features by: convolving multi-scale features with filters to obtain low-frequency components and multiple high-frequency components; and then transposing and convolving the low-frequency components and multiple high-frequency components to obtain wavelet enhancement features.

[0011] A further technical solution is to add the high-frequency components in the wavelet enhancement features to obtain the boundary features.

[0012] A further technical solution yields the following multi-scale enhanced features: Calculate the cosine similarity between multi-scale features and each prototype in the polyp prototype library, boundary prototype library, and background prototype library, and normalize the attention weights based on the cosine similarity. Based on the maximum value of the attention weight, the information entropy is calculated, and the uncertainty weight is obtained after normalization. Multi-scale enhanced features are obtained by fusing multi-scale features and reconstructed features based on uncertainty weights.

[0013] Further technical solutions include using cross-entropy, Dice loss, and KL divergence as losses for polyp regions, and using Dice loss as a loss for boundaries.

[0014] Secondly, the present invention provides a colon polyp segmentation system based on prototype memory and feature enhancement, comprising: The data acquisition module is configured to acquire the colonoscopy image to be segmented and preprocess it to obtain an edge feature map; The model segmentation module is configured to input edge feature maps into the image segmentation model for processing to obtain segmentation results. The image segmentation model comprises an input layer, a multi-branch feature enhancement and fusion module, and parallel segmentation heads and boundary heads, all connected in sequence. The multi-branch feature enhancement and fusion module includes multiple branches, each of which comprises a backbone network, a receptive field module, a discrete wavelet transform module, a semantic prototype memory module, and a decoder module, all connected in sequence. It is constructed as follows: The edge feature map is fed into the multi-branch feature enhancement and fusion module to obtain the final fused feature; The final fused features are fed into the segmentation head and the boundary head to obtain the segmentation result.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the colon polyp segmentation method based on prototype memory and feature enhancement as described in the first aspect.

[0016] Fourthly, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the colon polyp segmentation method based on prototype memory and feature enhancement as described in the first aspect.

[0017] The above one or more technical solutions have the following beneficial effects: To address the challenge of segmenting weakly bounded targets in medical images, this invention proposes a novel solution that integrates dynamic class prototype memory and boundary enhancement mechanisms, effectively improving the segmentation accuracy and model generalization ability of colon polyps.

[0018] This invention solves the problems of target ambiguity, morphological variability and noise interference in medical image segmentation by working together with two innovative features: feature adaptive fusion with uncertainty perception and dynamic prototype memory enhancement. It has achieved significant progress in segmentation accuracy, robustness and generalization.

[0019] The use of a dynamic class prototype memory enhances the image segmentation model's judgment confidence, improves its predictive ability, and makes the prediction boundaries clearer. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 This is a framework structure diagram of the image segmentation model according to an embodiment of the present invention; Figure 2 These are comparative visualizations of some image segmentation results from embodiments of the present invention. Detailed Implementation

[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0023] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0024] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0025] Terminology Explanation: Uncertainty estimation: The model automatically detects regions in the image that are difficult to determine (such as blurred edges) and labels them with confidence levels. Regions with low confidence levels will be preferentially referenced from the memory bank's experience.

[0026] The memory database stores an experience base of typical case features, continuously accumulating data on the size, texture, and edge characteristics of different polyps. When encountering a new case, it intelligently matches similar historical cases to assist in the judgment.

[0027] Class prototype: A "standard feature template" for a certain type of lesion, such as the shape and color patterns of common polyps. The model quickly identifies the type of lesion by comparing the similarity between the features of a new image and the class prototype.

[0028] Example 1 This embodiment discloses a method for segmenting colonic polyps based on prototype memory and feature enhancement. The method includes the following steps: S1: Obtain the colonoscopy image to be segmented and preprocess it to obtain the edge feature map; In this embodiment, the Kvasir-SEG dataset is used. It is an authoritative benchmark in the field of colon polyp segmentation, jointly constructed by the University Hospital of Oslo and the Simela Research Laboratory. This dataset provides crucial support for the development of deep learning algorithms through a standardized annotation process. As a specialized extension of the Kvasir series of datasets, its initial purpose was to advance the development of early colorectal cancer screening technologies, particularly the iterative optimization of computer-aided diagnostic systems.

[0029] The dataset contains 1000 clinically validated colonoscopy images and their pixel-level annotations. All samples are selected from examination videos of real-world diagnostic and treatment scenarios. Each image underwent multiple rounds of annotation review by a team of gastroenterology experts to ensure the medical accuracy of lesion boundaries. The sample features cover multiple dimensions of clinical challenges, specifically: 8 types of polyp morphological differences, including flat and pedunculated polyps; 12 types of endoscopic environmental interferences, such as mucosal folds obscuring the surface and instrument reflections; and a resolution range spanning from 720×576 to 1920×1072 pixels, fully simulating equipment differences in actual diagnosis and treatment. Notably, the dataset contains 12.3% of tiny polyp samples (less than 5 mm in diameter), posing a significant challenge to the algorithm's fine-grained recognition capabilities in complex anatomical contexts.

[0030] This dataset has become a benchmark platform for mainstream segmentation models such as U-Net and DeepLab, and its clinical value is reflected in two dimensions. From the perspective of algorithm development, it supports the construction of prototype intelligent endoscopy systems with real-time processing capabilities. At the clinical validation level, the complex cases in the dataset provide stress testing scenarios for evaluating the robustness of the algorithms, especially in terms of the segmentation stability in areas with artifact interference and low contrast, which has unique validation value.

[0031] The colonoscopy images were preprocessed using the Canny edge detection algorithm from the OpenCV library to extract edge features from the medical mask, resulting in an edge feature map. The processing mainly consisted of three key stages: First, the original binary mask was converted to grayscale to eliminate color channel interference and fully preserve anatomical structural information within the 0-255 pixel value range. Then, the classic Canny operator was used to perform gradient calculation, with high and low thresholds set to 100 and 200 respectively for non-maximum suppression, accurately capturing the transition region between polyp tissue and healthy mucosa. Finally, the detection results were standardized, generating a 0 / 1 binary matrix, and then linearly transformed by pixel values ​​to obtain a PNG format output conforming to medical imaging standards.

[0032] S2: Input the edge feature map into the image segmentation model for processing to obtain the segmentation result.

[0033] In this embodiment, the image segmentation model of this application is designed to address technical pain points such as high foreground-background similarity and blurred boundaries in colonoscopy images.

[0034] like Figure 1 As shown, the image segmentation model includes an input layer (STEM), a multi-branch feature enhancement and fusion module, a segmentation head, and a boundary head connected in sequence. The multi-branch feature enhancement and fusion module includes multiple branches, each of which includes a backbone network, a receptive field module (RFB), a discrete wavelet transform module (DWT), a semantic prototype memory module, and a decoder module connected in sequence.

[0035] The input layer contains convolutional layers and max-pooling layers. 7×7 convolutions are preferred for compressing semantic information, making the information more abstract and reducing computation. The receptive field (RFB) module includes multi-branch convolutional structures and dilated convolutions. The multi-branch convolutional structure comprises multiple parallel convolutional layers using kernels of different sizes (e.g., 1×1, 3×3, 5×5) to capture multi-scale features. The results of multiple convolutional layers are concatenated at the channel layer and then dilated convolutions are used to cover a larger image area, effectively capturing contextual information.

[0036] The image segmentation model extracts multi-scale features using a ResNet50 backbone network. , , , The multi-scale features are fed into the Receptive Field Block for preprocessing, and then into the Discrete Wavelet Transform module for further processing. The output of the Discrete Wavelet Transform module consists of two parts: one part is the wavelet enhancement features. , , , The other part is boundary features; these boundary features are passed to the previous layer module (i.e., a layer larger than the current layer). First, the boundary features are upsampled using the difference, and then added to the output features of the previous module to enhance the boundary features. Wavelet enhancement features are passed to the Semantic Prototype Memory (SPM) module, which uses the prototype enhancement features to obtain... , , , These features are fed into the decoder for multi-scale feature fusion. The features output by multiple decoders are passed through a convolutional layer (preferably 1×1 convolution) and a sigmoid activation function, and finally fed into the segmentation head and boundary head to obtain the final prediction result.

[0037] (a) Semantic Prototype Memory Module The Semantic Prototype Memory Module (SPM) is a dynamic feature enhancement mechanism designed for medical image segmentation tasks. This module effectively addresses issues such as variable target morphology, blurred boundaries, and noise interference in medical images by constructing a multi-class prototype memory and combining it with an uncertainty-driven adaptive fusion strategy. The SPM in this invention consists of three branches, each processing the semantic features of polyps. Boundary geometric features and background semantic features During model training, the input features for SPM are labeled, including polyp semantic feature labels, boundary geometric feature labels, and background semantic feature labels. Considering the possibility of incorrect labels, this invention assumes a set probability (preferably 0.75) for the labels, otherwise updates the prototype memory based on semantic similarity. During model prediction, similarity is calculated, and the most similar prototype is used as the current feature.

[0038] The construction process of a multi-category prototype memory is as follows: The proposed prototype memory It contains Q prototypes, for each prototype ( The first one representing the prototype memory bank (Prototype) and Input Feature Map The dimensions are the same, where C is the number of channels, H and W are the spatial dimensions, and i and j are the position indices.

[0039] For each feature Calculate its cosine similarity with each prototype, expressed as:

[0040] in, Indicates the first Three prototypes are constructed. The similarity between each feature and the prototypes in each of these three prototype libraries is calculated. Based on the calculated cosine similarity, select features... Most similar prototype index , is represented as:

[0041] Each prototype is updated using the exponential moving average method based on the prototype index, as follows:

[0042] in, Indicates the attenuation coefficient. This indicates an indicator function, where B is the batch size.

[0043] Now, features are reconstructed through the representation of the prototype. Reconstruction features The set is Specifically: For the obtained cosine similarity Soft attention is obtained using the softmax method:

[0044] Where Q represents the number of prototypes. The prototype number is Q; there are a total of Q prototypes in a prototype memory. This is the soft attention value. Features With the Cosine similarity of the prototypes Features With the Cosine similarity of the prototypes.

[0045] Then, the Gumbel-Max technique is used to obtain the unique hot attention, represented as:

[0046] in, For unique heat-focused attention value, This is an operation to convert data into one-hot encoded data. , This is random noise sampled from the Gumbel-Softmax distribution.

[0047] Then obtain differentiable attention , is represented as:

[0048] in, This is the temperature coefficient.

[0049] Finally, calculate hard attention. , is represented as:

[0050] in, To stop the gradient operation, finally, enhanced features are obtained by aggregating all prototypes. .

[0051] Use the methods described above to construct polyp prototype libraries respectively. Boundary Prototype Library and background prototype library ,in , , These represent the number of prototypes in the corresponding prototype library.

[0052] Based on the constructed polyp prototype library, boundary prototype library, and background prototype library, calculate the cosine similarity between the feature and each prototype in each prototype library:

[0053] Then normalize the attention weights:

[0054] To estimate uncertainty, this invention uses the maximum value of each class as the probability of the class, and then calculates the information entropy. :

[0055] in, For the first The similarity matrix between the class prototype and the current feature, with size Q; For constant terms, The number of classes (the number of prototype memories). The ordinal number of the current class (the first one) (A prototype memory bank).

[0056] Then, normalize the entropy using the maximum and minimum values:

[0057] The enhanced features obtained by fusing the original features and the reconstructed features using uncertainty weights are represented as follows: ,

[0058] in, For Hadama accumulation.

[0059] (ii) Discrete wavelet feature enhancement Because polyps are highly similar to the background, identifying lesion areas is difficult. To alleviate this problem, this invention proposes using discrete wavelet transform to enhance features. First, a two-dimensional Haar discrete wavelet transform is implemented using convolution operations in two feature dimensions, with a stride of 2. The four filters are as follows:

[0060] in, It's a low-pass filter, the other three ( , , Let be a high-pass filter. For each input channel, it has four output channels, and the size is half the original size, represented as:

[0061] in, Original features The low-frequency components, These are the vertical, horizontal, and diagonal high-frequency components, respectively.

[0062] Wavelet-enhanced feature information is obtained through transposed convolution. :

[0063] Multiscale features , , , After discrete wavelet feature enhancement, the following were obtained: That is, wavelet enhancement features , , , .

[0064] in, It is applied through a learnable convolution operation. What was obtained , .

[0065] Sum all high-frequency components as boundary features Boundary features will be added to the features of the previous level to enhance the boundary information, as described below: .

[0066] (III) Loss Function For each layer of features, the segmentation head and boundary head are used to obtain the predicted polyp regions respectively. With boundary Then use real labels. With boundary Calculate the corresponding loss to train the model. Segmentation loss. Cross-entropy, Dice loss, KL divergence, and boundary loss are used. This is a loss for Dice.

[0067] For polyp regions, cross-entropy, Dice loss, and KL divergence are used as losses; for boundaries, Dice loss is used as the loss.

[0068] The expressions for the binary cross-entropy loss, Dice loss, and KL divergence are as follows:

[0069]

[0070]

[0071] in, The number of pixels; This is a label with a value of 0 or 1; The result predicted by the model has a value of [0,1]. This is a constant term to prevent division by zero.

[0072] The overall loss function expression is: .

[0073] like Figure 2 As shown, experiments are conducted to provide a visual comparison of some image segmentation results. Through the above technical features, the prototype memory records information from all training sets, and uncertainty estimation is combined to enhance judgment confidence. Feature storage and utilization capabilities are proposed, and boundary detection is optimized.

[0074] In summary, this invention proposes a deep learning framework. First, the input colonoscopy image is preprocessed to extract edge features. The core of the image segmentation model adopts an encoder-decoder structure. In the encoding stage, multi-scale features are extracted through a ResNet50 backbone network and then processed sequentially through a receptive field module and a multi-branch discrete wavelet transform module to enhance contextual information and separation boundary features. Subsequently, an innovative semantic prototype memory module is introduced. This module utilizes three dynamically updated prototype libraries—polyps, boundaries, and background—and adaptively enhances semantic features by calculating uncertainty weights. In the decoding stage, the enhanced features from each scale are fused, and the prediction results are finally output through parallel segmentation heads and boundary heads. The overall training is supervised by a composite loss function. This method aims to improve the model's segmentation accuracy and robustness for polyps with variable morphology and blurred boundaries through uncertainty perception and memory enhancement mechanisms.

[0075] Example 2 This embodiment discloses a colon polyp segmentation system based on prototype memory and feature enhancement, including: The data acquisition module is configured to acquire the colonoscopy image to be segmented and preprocess it to obtain an edge feature map; The model segmentation module is configured to input edge feature maps into the image segmentation model for processing to obtain segmentation results. The image segmentation model comprises an input layer, a multi-branch feature enhancement and fusion module, and parallel segmentation heads and boundary heads, all connected in sequence. The multi-branch feature enhancement and fusion module includes multiple branches, each of which comprises a backbone network, a receptive field module, a discrete wavelet transform module, a semantic prototype memory module, and a decoder module, all connected in sequence. It is constructed as follows: The edge feature map is fed into the multi-branch feature enhancement and fusion module to obtain the final fused feature; The final fused features are fed into the segmentation head and the boundary head to obtain the segmentation result.

[0076] Example 3 The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method of Embodiment 1.

[0077] Example 4 The purpose of this embodiment is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the method of Embodiment 1.

[0078] The steps and methods involved in the apparatuses of Embodiments 3 and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0079] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0081] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for segmenting colonic polyps based on prototype memory and feature enhancement, characterized in that, include: The colonoscopy image to be segmented is acquired and preprocessed to obtain the edge feature map; The edge feature map is input into the image segmentation model for processing to obtain the segmentation result; The image segmentation model comprises an input layer, a multi-branch feature enhancement and fusion module, and parallel segmentation heads and boundary heads, all connected in sequence. The multi-branch feature enhancement and fusion module includes multiple branches, each of which comprises a backbone network, a receptive field module, a discrete wavelet transform module, a semantic prototype memory module, and a decoder module, all connected in sequence. It is constructed as follows: The edge feature map is fed into the multi-branch feature enhancement and fusion module to obtain the final fused feature; The final fused features are fed into the segmentation head and the boundary head to obtain the segmentation result.

2. The colon polyp segmentation method based on prototype memory and feature enhancement as described in claim 1, characterized in that, The colonoscopy images are preprocessed using an edge detection algorithm to obtain edge feature maps.

3. The colon polyp segmentation method based on prototype memory and feature enhancement as described in claim 1, characterized in that, The multi-branch feature enhancement and fusion module is constructed as follows: The edge feature map is fed into the multi-branch backbone network for feature extraction to obtain multi-scale features, which are then fed into the receptive field module. After preprocessing the multi-scale features in the receptive field module, the features are fed into the discrete wavelet transform module for further processing to obtain wavelet enhancement features and boundary features. The wavelet enhancement features are fed into the semantic prototype memory module, combined with the boundary features of the next level scale for enhancement, to obtain multi-scale enhancement features, which are then fed into the decoder. The decoder fuses the multi-scale enhanced features to obtain the final fused features.

4. The colon polyp segmentation method based on prototype memory and feature enhancement as described in claim 3, characterized in that, The wavelet enhancement features are obtained by convolving the multi-scale features with a filter to obtain low-frequency components and multiple high-frequency components; and by transposing and convolving the low-frequency components and multiple high-frequency components to obtain wavelet enhancement features.

5. The colon polyp segmentation method based on prototype memory and feature enhancement as described in claim 4, characterized in that, The boundary features are obtained by adding the high-frequency components in the wavelet enhancement features.

6. The colon polyp segmentation method based on prototype memory and feature enhancement as described in claim 3, characterized in that, The multi-scale enhancement features obtained are as follows: Calculate the cosine similarity between multi-scale features and each prototype in the polyp prototype library, boundary prototype library, and background prototype library, and normalize the attention weights based on the cosine similarity. Based on the maximum value of the attention weight, the information entropy is calculated, and the uncertainty weight is obtained after normalization. Multi-scale enhanced features are obtained by fusing multi-scale features and reconstructed features based on uncertainty weights.

7. The colon polyp segmentation method based on prototype memory and feature enhancement as described in claim 1, characterized in that, For polyp regions, cross-entropy, Dice loss, and KL divergence are used as losses; for boundaries, Dice loss is used as the loss.

8. A colon polyp segmentation system based on prototype memory and feature enhancement, characterized in that, include: The data acquisition module is configured to acquire the colonoscopy image to be segmented and preprocess it to obtain an edge feature map; The model segmentation module is configured to input edge feature maps into the image segmentation model for processing to obtain segmentation results. The image segmentation model comprises an input layer, a multi-branch feature enhancement and fusion module, and parallel segmentation heads and boundary heads, all connected in sequence. The multi-branch feature enhancement and fusion module includes multiple branches, each of which comprises a backbone network, a receptive field module, a discrete wavelet transform module, a semantic prototype memory module, and a decoder module, all connected in sequence. It is constructed as follows: The edge feature map is fed into the multi-branch feature enhancement and fusion module to obtain the final fused feature; The final fused features are fed into the segmentation head and the boundary head to obtain the segmentation result.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the colon polyp segmentation method based on prototype memory and feature enhancement as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the colon polyp segmentation method based on prototype memory and feature enhancement as described in any one of claims 1-7.

Images