Research And Application On Medical Image Recognition Algorithm Via Deep Learning

Recently,Deep Learning(DL),which can learn highly rep-resentative and hierarchical features of medical images from data-driven training process,has become the primary method for medical image anal-ysis task.In this thesis,we will lucubrate the DL-based recognition al-gorithms suitable for the detection of liver lesions and the diagnosis of esophageal contraction vigor,and the main research contents and innova-tive achievements are as follows:The hyperplasia of hepatic nodules,the accumulation of hepatic as-cites,and hepatomegaly are indicative of liver with organic lesions.We creatively presented an end-to-end Computer Aided Diagnosis(CAD)sys-tem for recognition,acquisition,segmentation and quantification of liver lesions on large-scale Computed Tomography(CT)sequences.This sys-tem is a three-task processor:(1)Acquisition Conv Net,which is similar in structure to VGG Net but smaller in scale,selects abdominal CT images containing liver lesions from a large number of CT slices.(2)In order to quickly locate and accurately segment fibrotic nodules,Detection Conv Net was designed to extract and combine multi-scale features to refine the seg-mentation boundary through dense dilated convolution pyramid pooling.(3)Segmentation Conv Net is composed of multiple homogeneous Res C Units that combine residual learning and channel combination to capture local context features by down-sampling,and the PD Unit that is able to ex-tract global semantic feature with multi-scale information,which achieves the pixel-based classification task of hepatic ascites and liver.A lots of em-pirical studies have shown that our CAD framework has strong detection performance and application potential in rapid acquisition of interested CT slices,accurate localization and complete dissection of liver lesions.High-Resolution Manometry(HRM)image is a kind of novel medical image for examinating esophageal motility health,which has been widely used in clinical trials.However the diversity and the complexity diagno-sis of Esophageal Motility Disorder(EMD)lead to frequent misdiagnosis.To improve the accuracy of clinic diagnosis,a Convolutional Neural Net-work(CNN)based esophageal contraction vigor classifier(Po S-Clas Net)is designed to assist doctors in checking out esophageal body health status.As a multi-task model,Po S-Clas Net made up of Po SNet and S-Clas Net.The former is an encoder-decoder model,which detects and extracts the swallowing frame in HRM images that records changes in esophageal con-traction pressure during swallowing,and the latter is a lightweight classifi-cation model based on Smaller VGG,which learns the features of swallow-ing pressure to correctly diagnose the type of contraction vigor.In clinical data sets,Po S-Clas Net shows better comprehensive performance than other classification models,which reflects an objective fact of its great adaptation to HRM images features.