Research On Colorectal Polyp Image Segmentation Methods Based On Deep Learning

In the field of medical imaging,image segmentation technology plays a vital role in processing various types of images such as CT,MRI,X-rays,ultrasounds,and endoscopic images.Among these,polyp segmentation under endoscopy is particularly crucial,as doctors need to identify potential colorectal polyps during examinations,which is essential for the early diagnosis and treatment of colorectal cancer.However,endoscopic image segmentation faces multiple challenges such as blurred image boundaries,large scale variation,uneven lighting,and low contrast.These factors make the screening of colorectal polyps both time-consuming and reliant on professional expertise.Consequently,computer vision-based polyp segmentation methods have shown great potential.They can improve the efficiency of polyp diagnosis and reduce misdiagnoses and missed diagnoses.This dissertation directly addresses these challenges.It thoroughly analyzes the strengths and weaknesses of existing methods and discusses how to apply advanced deep learning technologies,including feature enhancement and feature decoupling strategies,to propose a series of innovative solutions aimed at improving the accuracy and efficiency of polyp segmentation in endoscopic images.The main content and achievements of this dissertation are as follows:1.To address the issue of inaccurate polyp edge segmentation,a polyp segmentation network with feature alignment and boundary optimization was designed.This architecture adopts a new paradigm of feature fusion,using a learned semantic offset field to align multi-level features,significantly improving the accuracy of polyp edge localization.In addition,an auxiliary boundary branch was designed to enhance the model’s perception of polyp boundaries,thereby improving the performance of polyp boundary prediction.After end-to-end optimization,the reference boundary map is used as a supplement to high-level semantic representations,further improving the effectiveness and efficiency of polyp segmentation.2.To address missegmentation and incorrect segmentation of polyps due to uneven lighting and low contrast in endoscopic images,a Feature Augmentation Pyramid Network(FAPN)was proposed.FAPN uses a cross-embedding module to efficiently extract and merge polyp features,and a prediction calibration module to precisely locate polyp features,effectively highlighting areas of interest and significantly reducing the impact of uneven lighting and low contrast.Additionally,its hierarchical feature fusion module produces a more robust multi-scale polyp feature representation,further enhancing the accuracy of polyp segmentation.Experimental validation showed that FAPN outperformed existing advanced methods on multiple polyp segmentation datasets.3.To overcome limitations in feature propagation and fusion efficiency in current polyp segmentation technologies,an innovative Rectified Feature Pyramid Network was proposed.The core goal of this network is to optimize and correct the feature aggregation and propagation mechanism within the polyp segmentation framework,achieving more efficient and accurate feature processing.By introducing an efficient feature propagation enhancement module,the network achieves precise transfer of full-scale features at all stages.Additionally,its unique gating mechanism effectively strengthens feature fusion,improving the accuracy of information filtering and processing.4.To address the shortcomings of existing polyp segmentation methods in simultaneously enhancing the interior consistency of polyps and optimizing edge performance,a comprehensive optimization method for polyp features,the Feature Decoupled Network(Fe DNet),was proposed.Fe DNet applies feature decoupling techniques from the Laplacian pyramid to effectively decompose and finely optimize polyp features,thereby enhancing baseline performance and significantly surpassing existing state-of-the-art methods,demonstrating its excellent generalization ability on multiple datasets.5.To address the limitations of Fe DNet in distinguishing high and low frequency features of polyp body and edge features,an Adaptive Feature Decoupled Module(AFDM)was proposed.AFDM combines Fourier transform and learnable masks to effectively decouple high and low-frequency features in the frequency domain,enhancing the accuracy of feature decoupling.This improvement enables the feature decoupled network integrated with AFDM to demonstrate superior performance and generalization ability on multiple datasets.In summary,the main work of this dissertation is to design and implement a series of colorectal polyp segmentation methods,aimed at enhancing the accuracy and efficiency of colorectal polyp segmentation in endoscopic images.These methods have demonstrated outstanding performance on multiple polyp segmentation datasets and exhibit excellent generalization capabilities.This research opens new horizons for the early diagnosis and treatment of colorectal cancer,with significant clinical significance.