Research On Three-dimensional Reconstruction And Computer-aided Diagnosis Algorithm Of Medical Image

According to the World Health Organization,lung cancer has become the cancer with the highest number of confirmed cases and death toll.The use of computer technology to screen the early form of lung cancer(i.e.pulmonary nodules)is conducive to early detection and treatment,thus improving the survival chance of lung cancer,reducing the burden of reading films,and assisting doctors in diagnosis.In recent years,with the development of deep learning technology,it has also been applied to various fields including medical image analysis.The detection and segmentation of pulmonary nodules using deep learning is one of the hot research topics in computer-aided diagnosis.Based on the chest CT data,this thesis aims at exploring the utilization of deep learning and three-dimensional reconstruction technology to provide radiologists with a more accurate and effective assisted diagnosis way of easily observing pulmonary nodules.The main contents of this thesis are as follows:Firstly,the method of pulmonary nodules detection and segmentation based on Mask R-CNN was studied.Since medical image has the characteristics of small sample size and unbalanced positive and negative samples,this thesis adopts the Mask R-CNN with the backbone transfer learning mechanism,two-stage structure to balance positive and negative samples,and simultaneously detection and segmentation function in the same network.Besides,the influence of loss functions with different weights on multitask was explored.At present,most studies are concentrated on the detection of pulmonary nodules and lack of publicly availabe dataset for the segmentation task.Therefore,this thesis preprocessed the benchmark dataset of the detection of pulmonary nodules,LUNA16 dataset.Then 2763 axial-plane view dataset with contour labels of nodules was constructed using labelme tool based on the diameter of pulmonary nodules marked by radiologists.The new dataset was called labelme＿LUNA16.Experiments were carried out on LUNA16,labelme＿LUNA16 and Ali Tian Chi dataset.As regards the detection of pulmonary nodules,on the labelme＿LUNA16 dataset,the sensitivity of 88.1% at 1 FP/scan and CPM score of 0.796 were gotten,which outpeforms most recent studies.Meanwhile,the model was evaluated on independent Ali Tian Chi dataset and the CPM score of 0.625 was obtained,which indicated good generalization and robustness.Furthermore,with respect to segmentation of pulmonary nodules,AP@50 score of 0.882 on labelme＿LUNA16 databaset was obtained.Secondly,the three-dimensional reconstruction methods of medical image were invistigated.This thesis focused on marching cubes algorithm and ray-casting rendering algorithm which are representative algorithm of suface rendering and volume rendering,respcetively.The implementations,advantages and disadvantages of these two algorithms were analyzed.Comparative experiments were carried out on two algorithms with the dental CT data to analyze the rendering effect of different parameters and algorithms.Then a three-dimensional reconstruction experiment of lung and pulmonary nodules was conducted.Finally,a three-dimensional visualization auxiliary diagnosis system for pulmonary nodules was designed based on MITK platform.The three-dimensional reconstruction results of pulmonary nodules were consistent with the position and size of the axial-plane view,sagittal-plane view and coronal-plane view.And the three-dimensional model could also be easily interacted which shows the effectiveness and operability of the system.