Research On Medical Image Segmentation Algorithms Based On Deep Learning

Medical image segmentation is one of the important steps in clinical diagnosis.Accurate segmentation of lesions or tumors is one of the important bases for killing pathological cells and protecting non-pathological areas in clinical treatment.In practical application,the important step of segmentation is usually carried out manually by a professional doctor with rich practical experience,which takes a lot of work and time.Introducing computer-aided diagnosis system to automatically segment regions of interest in medical images can greatly improve the efficiency of doctors.Traditional automatic segmentation methods are mostly based on image processing technology,and rely on some shallow features,such as gray level,texture and so on to achieve segmentation.These shallow features have poor robustness and anti-noise ability,and can not adequately express the intrinsic characteristics of the region of interest,so the generalization ability of the model is not good.Deep learning technology can automatically learn deep abstract features with good discriminant and generalization ability from data.It can improve the accuracy of segmentation results by applying it to medical image segmentation.The main contents of this thesis are as follows:(1)A medical image segmentation framework based on depth learning is proposed.Based on Fully Connection Net(FCN),this thesis modifies the convolution layers to 3D convolution,and adds skip-connection to the network to make it conform to the architecture of U-Net network,which is used to segment lung nodes in CT images.(2)The method of reducing false positives in segmentation results is studied.Ablation experiment was carried out to extract lung regions and resample CT images in data preprocessing to observe the effect of false positive segmentation results.(3)The effectiveness of skip-connection in U-Net network is discussed.In this thesis,skip-connection in U-Net is tested by Ablation test,and the effects of skip-connection on recall rate and false positive rate of segmentation results are compared.(4)A false positive attenuation algorithm is proposed.In this thesis,a false positive attenuation test is carried out on the segmentation results.The specific method is to train a classification network,further screen the structure of the segmentation,judge whether the segmentation results are real nodes,and draw the FROC curve of the corresponding results.This method was evaluated on the test set A of Tianchi Medical AI Competition.The recall rate of segmentation reached 92%,and the average false positive rate of each CT was only 10%.In order to compare with other methods,we also evaluated on the public data set LUNA16.The recall rate of the final segmentation is 93%.From the final FROC curve,we can see that our method is superior to most of the algorithms on the LUNA 16 list in recall rate.