Reseach On Medical Image Segmentation Technology Based On Deep Learning

Medical imaging is the main means of doctor’s clinical diagnosis.It can assist doctors to make more scientific diagnosis to provide an important basis for prevention and treatment.Medical image segmentation is mainly based on some similar features of the image is divided into several different regions,so as to obtain segmentation results.At present,the division of medical image has a great application prospect in visual diagnosis,the recognition and localization of the contour and edge of the lesion and the implementation of surgical plan.Based on this,this paper selected rectal polyps as experimental background in the field of medical imaging to study.Colorectal cancer is a kind of malignant gastrointestinal tumor,and also one of the main cancer components at present.Most colorectal cancer is evolved from adenomatous polyps.If early intervention can be carried out,colorectal cancer can be effectively prevented.Generally,the most effective screening and diagnosis method is colorectal endoscopy,but the accuracy of the examination is extremely dependent on the expertise of doctors and the diagnostic efficiency is low.Therefore,it is urgently needed to find a method to quickly and accurately locate polyps in colorectal endoscopy.To this end,this paper conducts research based on deep learning methods,and proposes two models for different needs: 1.FRCNet ensures relative accuracy while the number of parameters and GFLOPs are only 780,000 and 3.36%,respectively.Compared with the current training network models,which generally have hundreds of millions of parameters,FRCNet greatly reduces the computational effort and greatly reduces the hardware cost if it can be produced on a large scale subsequently.2.AGCNet has relatively high hardware requirements.After validation of the Kvasir-SEG dataset,its Dice coefficients and Precision reached 88.99% and91.01%,respectively,with accuracy surpassing all advanced models in the control group of this paper,while its FPS can be maintained at 33,taking into account the inherent real-time needs of clinical diagnosis.For future practical clinical applications,trade-offs can be made on a case-by-case basis.The following is a specific overview for both models.(1)A novel and lightweight context adaptive network(FRCNet: Feature Refining and Context-Guided Network for Efficient Polyp Segmentation),Based on U-NET,FRCNet developed three modules embedded into U-NET,which can be used for real-time polyp segmentation efficiently.In order to alleviate the interference from background noise of polyp image and efficiently identify polyps from the background,this paper first adopted an enhanced context self-correction module(ECC: Enhanced Context Calibrated Module),which can model spatial dependencies at a distance through Context correction operations,thereby enabling more discriminative features to be proposed.In addition,this paper designed a Progressive context-aware Fusion module(PCF: Progressive Context-aware Fusion module)to dynamically capture polyps of different scales by collecting multi-range Context information.Finally,a multi-scale pyramid aggregation(MPA: multi-scale pyramid aggregation)module is designed.The significance of developing this module is to learn richer representative features and integrate segmentation results for refining networks.The above three modules are integrated to form FRCNet in this paper.Meanwhile,a large number of experiments on Kvasir-SEG and CVC-Clinic DB data confirm the effectiveness of the network model proposed in this paper,which exceeds some advanced network models at present.(2)A precise adaptive global context network(AGCNet: a Precise Adaptive Global Context Network for Real-time Colonoscopy)is proposed.Firstly,in order to adapt to the problem of large-scale variation of polyps,we design a granular-level and multi-scale semantic fusion module(MSFM: Multi-scale Semantic Fusion Module),which is based on multiple filters to collect more easily neglected detailed features to improve the characterization ability of the network,so as to achieve accurate segmentation of polyps with variable scales.If we simply use complex spatial pixels to model the long-range dependence,it is not only easy to introduce more background noise and increase the computational effort.Therefore,this paper designs a pyramid context aggregation module(CPAM: Context-aware Pyramid Aggregation Module),which includes a dual-attention mechanism.The CPAM guides the aggregation of features in different regions to increase the confidence of the network with rich semantic information,and then further reinforces the important features and suppresses the features in non-target regions through the dual-attention mechanism.The seamless integration of the above two approaches form our proposed AGCNet,which is also validated on the Kvasir-SEG and CVC-Clinic DB datasets,confirming the effectiveness of our model.In this paper,the attentional mechanism is carried through our proposed FRCNet and AGCNet,and also reflects the importance of the attentional mechanism for medical image segmentation;the research in this paper has important implications for future clinical diagnosis,and complementary medicine.