Colorectal Polyp Image Segmentation Based On Deep Learning

Colorectal cancer is the third most prevalent cancer at present.Colon polyps are characteristic of its early lesions,and early diagnosis is an important means of preventing colorectal cancer;studies have shown that regular colonoscopy and early prevention by judging the location and size of colon polyps can reduce the incidence of colorectal cancer by30%.In clinical practice,the discovery of suspicious polyp-like lesions requires repeated close observation and confirmation by pumping and flushing.Large polyps are usually not missed,but very small or hidden polyps in the corners are inevitably missed by the doctors doing colonoscopy,even if they are famous and carefully observed,and there is a certain probability of misdiagnosis and omission.Segmentation of colorectal polyp images helps to improve doctors’ full utilization of colonoscopic images to discover anatomical and case information reflected by polyp images,which facilitates doctors’ diagnosis and condition assessment;based on this,this paper constructs a deep network based on codec structure and incorporates Markov random field technique for colorectal polyp image segmentation.To address the characteristics of polyp images and the difficulties in segmentation,we propose a deep aggregation network based on Res Ne St-50 and a multi-branch aggregation network based on Transformer for polyp image segmentation,and build a Markov random field-based tree-weighted information transfer(TRW-S)algorithm in the post-processing of the network to solve the problems of segmentation of tiny lesions and blurred edge segmentation of polyp images.and other problems.The main research contents and specific work of this paper are reflected in the following aspects(1)To address the problems of existing technologies in processing polyp image segmentation and low accuracy of small lesion segmentation,a deep aggregation network based on Res Ne St-50 is constructed in this paper for polyp image segmentation algorithm.By introducing the split-attention mechanism at the coding side,the model encoder’s ability to express polyp image features is enhanced,the network’s ability to extract polyp image features is improved,and the segmentation effect of the model on small lesions of polyp images is improved.Meanwhile,the model in this paper adopts a combination of aggregation network and multi-scale structure,which effectively solves the problems of varying lesion size and segmentation difficulties,and improves the accuracy and robustness of the model polyp image segmentation more significantly.(2)To address the problems of poor internal coherence and unclear segmentation edges of polyp image segmentation,this paper proposes a tree weighted information transfer(TRWS)algorithm based on Markov random field for post-polyp image segmentation processing;by establishing Markov random field in the polyp image after network segmentation and solving the minimum value of its global energy,the problems of internal incoherence and unclear polyp image segmentation are solved.The problem of incoherent polyp image segmentation and unclear edge of polyp image segmentation is solved.The post-processing algorithm in this paper effectively improves the accuracy and robustness of polyp image segmentation,which facilitates the visualization and understanding of polyp images and enhances the accuracy of diagnosis.(3)To address the problems that the segmentation effect of the model decreases when the tissue scales of polyps differ greatly,and the poor generalization ability and slow segmentation speed in this case,this paper constructs a polyps image segmentation architecture based on PVTv2 multi-branch aggregation network by introducing the Transformer structure to achieve the extraction of global pixel relationships of polyps images.In addition,by adopting the multi-branch prunable structure in the model,the accuracy of model segmentation is improved while avoiding the problem of increasing the prediction time.(4)To verify the rationality of the proposed model structure,a large number of experiments are conducted in this paper,including the selection of different polyp image datasets for comparison experiments,comparison experiments with existing polyp image segmentation networks,comparison experiments of generalization ability,and ablation experiments.The experimental results on the open dataset of colorectal polyps Kvasir-SEG show that the polyp image segmentation based on Res Ne St-50 multi-scale aggregation network can reach 91.6% m Dice and 86.3% m Io U;the model shows good generalization ability when tested on unknown datasets ETIS-Larib Polyp DB,Colon DB,and compared with Compared with the benchmark SFA network,m Dice improved by 14.2% and 7.7%,respectively;from the performance of this model on ETIS-Larib Polyp DB dataset,the Res Ne St-50 multi-scale aggregation network based on Res Ne St-50 is very sensitive to micro polyp lesion tissue,and the segmentation effect can be improved by 0.6% after postprocessing using TRW-S algorithm;for ETIS-Larib Polyp DB public dataset experimental results show that the polyp image segmentation based on PVTv2 multibranch aggregation network improves m Dice by 16.3% and m Io U by 14.8%,and the network has good generalization ability.The experimental results not only reflect the rationality of the polyp image segmentation model proposed in this paper,but also verify the effectiveness of this paper’s model for polyp image segmentation.