Research And Application Of Key Technology For Intelligent Classification Of Gastrointestinal Images

According to the World Health Organization(WHO),about 1.8 million people die directly or indirectly from gastrointestinal diseases worldwide each year.However,there is a problem of doctor-patient conflict caused by the small number of professional endoscopic analysts and the large amount of endoscopic data generated by patients.Therefore,artificial intelligence "AI" is considered to assist in GI image diagnosis,and AI-based GI diagnostic technology faces problems such as large amount of data,numerous classification network models,and data not shared among medical institutions.Therefore,this thesis addresses the aforementioned key issues.This thesis improve the network model selection method to address the problem of difficult network selection,and use deep reinforcement learning to build an intelligent selection framework for classification network models,avoiding the exhaustive search of network models using grid search methods,and optimize the problem that reinforcement learning requires queuing for obtaining multiple sets of feedback values on a single machine,using a distributed method for The network selection method is also optimized for the problem that multiple sets of feedback values need to be queued on a single machine,and a distributed method is used to compute the feedback values in parallel,thus reducing the waiting time for network queuing.The thesis also proposed a method to improve the model accuracy due to the uneven distribution of data samples among medical institutions,and used an adaptive algorithm to dynamically optimize the aggregation function in federal learning.It is demonstrated that the distributed training framework based on federated learning requires significantly lower time overhead than the stand-alone training mode and does not suffer from a greater loss of model accuracy than the stand-alone training approach.Finally,this thesis developed an intelligent diagnosis system for GI images,which is capable of diagnosing GI images,storing patient cases,and distributing training,and completed the system testing for this system.This thesis investigates the difficulties in selecting algorithms for network models and data privacy and security problems of "AI" assisted GI image diagnosis,and propose the research solution to address the problems.