Medical CT Image Segmentation Based On Multi-View And Multi-Task Learning

Medical CT examination has become an indispensable assistant tool for disease screening,diagnosis and treatment,and is increasingly widely used.However,the large number of medical CT images,the non-structural characteristics of image data and the professionalism of medical image content,which are generated everyday,bring great challenges to the rapid and accurate diagnosis and treatment based on medical image interpretation.In recent years,the automatic analysis of medical images and computer-aided diagnosis and treatment technology have received great attention,for the automatic interpretation and analysis of medical data can improve the accuracy,convenience and ability to process large-scale data of medical diagnosis and treatment.Providing accurate image analysis technology for screening,diagnosis and surgical planning of major clinical diseases is an urgent fr ontier problem in the field of image analysis.China is a country with high incidence of hepatocellular carcinoma,and hepatocellular carcinoma is a kind of malignant tumors with high incidence and mortality.Therefore,the study of computer-assisted diagnosis and treatment of hepatocellular carcinoma has important application value.In this paper,for the purpose of computer-aided diagnosis and treatment of hepatocellular carcinoma,the intelligent analysis of liver CT images is studied by using deep convolution neural network.Specifically,we propose effective methods to solve some problems in intelligent preprocessing of CT images and automatic segmentation of liver and tumors.In the pre-processing stage,a multi-task parallel convolution regression network is designed to predict the orientation deviation parameters,which solves the shortcomings of traditional registration correction methods in large-angle deviation correction and achieves good results in both large-angle deviation and small-deviation cases;furthermore,in view of the small amount of medical image data,multi-task parallel convolution regression network is serialized to predict the orientation deviation parameters,and the solution of deviation parameters is divided into two steps.By solving the simpler first stage parameters,the solution of the second stage problem is facilitated and the final correction result is improved.In the stage of liver segmentation,on the basis of traditional mean fusion method of three-orientation segmentation,aiming at the inaccuracy of segmentation results in edges and some special shapes without using the local information of three-dimensional images in fusion,an adaptive fusion method are designed to meature the confidence of segmentation results from different perspectives in different localities,and then weighted fusion.Thus,the final segmentation and fusion results are improved.In the segmentation of liver tumors,a multi-task segmentation network using Correlation-Loss is designed to solve the problems of under-segmentation of liver and over-segmentation of tumors.By sharing the shallow network layers,the tasks can promote each other and generate the shallow features with stronger generalization faster.Furthermore,a task-related Correlation-Loss function describing the geometric relationship between tasks is designed.Experiments on ISBI-2017 liver tumor segmentation dataset show that the proposed method is effective.In the pre-processing orientation correction experiment,the angular lateral correction error has been greatly improved compared with the best result of the reference registration correction method;in the multi-view liver segmentation problem,the ASSD error of the segmentation result is only about half of the reference experimental result;in the multi-task liver and tumor segmentation using association loss,the visualization results show that the proposed method effectively solves the problem of over-segmentation of liver and under-segmentation of tumors in abdominal CT image organ segmentation.