Research And Implementation Of Automatic Segmentation Of Colonoscopic Polyp Images Based On Deep Learning

As one of the most common types of cancer,colorectal cancer poses a threat to people’s health.The most common cause of colorectal cancer is polyps.The use of colonoscopy to detect polyps in the early stages of the disease is essential for prevention and treatment of colorectal cancer.Due to the similarity of polyp color,texture and background and the variability of polyp size and shape,the rate of missed diagnosis is high with naked eye.Especially in remote areas of China,the lack of medical staff and equipment leads to shorter colonoscopy times and higher rates of missed polyps.Therefore,the computer-aided diagnosis system is important for the precise localization and segmentation of polyps for the clinical treatment of physicians.In this study,we analyze the image features and segmentation difficulties of polyps,and conduct an in-depth study on polyp segmentation based on deep learning.Experiments are conducted on five public datasets,Kvasir,CVC-Clinic DB,CVCColon DB,ETIS and CVC-T,to validate the effectiveness of the proposed method,The details of the research are as follows:A multiscale and attention mechanism based model(FARes2Net)is proposed.The method uses the Res2 Net backbone as an encoder.Multi-scale convolution inside the feature layer forms different receptive field and obtain different fine-grained features.To address the problem of polyp size discrepancy,the model designs a pyramidal decoder in the decoder part,and uses the pyramid pooling module to fuse multi-scale features.To address the problem of unclear boundaries between polyp and background,an fusion attention module combining shallow attention and axial reverse attention mechanisms with attention gates is proposed.By using channel weighting to combine attention feature maps,the redundant information is reduced while the effective information is enhanced.The segmentation ability of the model for polyp boundaries is improved.In addition,a compound loss is proposed as the loss function of this algorithm to make the model more focused on the pixels at the polyp edge.By conducting a large number of comparison experiments,the proposed method enriches the extraction of detail information,and achieves accurate automatic polyp segmentation with m Dice and m Io U of 92.1% and 87.8% on the CVC-Clinic DB dataset.An adaptive threshold reflection detection and repair method is proposed.The method addresses the problem that highlighted areas in the colonoscopic images of the dataset can affect the segmentation accuracy,and reduces the impact of reflective areas on the segmentation results.In addition,a color exchange strategy is used to reduce the difference in color distribution of polyp images acquired under different lighting conditions and angles,which reduces the overfitting problem due to color and makes the model more focused on the effective features of the target polyps.A lightweight deep neural network(LMformer)based on the Mi T backbone is proposed.Compared to convolutional neural network encoders that lack context dependency,Mi T encoders can obtain global contextual information.A channel-wise feature pyramid module is used in the decoder part to obtain multi-scale information,which increases the receptive field of the model while maintaining lightweight.To extract detailed features of polyp edges,the channel spatial aggregation module is proposed in this thesis.The channel spatial aggregation module maintains a high internal resolution in the computation of channel and spatial attention while minimizing the dimensionality of the input feature maps.Through experiments,the network maintains light weight and efficiency while obtaining excellent segmentation results.Based on the trained polyp segmentation model,a simple web-based polyp segmentation system is designed and implemented,and the system modules are introduced through specific interfaces to verify the feasibility of applying this research technique to computer-aided diagnosis systems.