A Research On The Automatic Planning For Lung Cancer Radiotherapy Based On Deep Learning

Both morbidity and mortality of lung cancer continue to be very high,which is a serious threat to human's health.As one of the three major treatments for cancer,radiotherapy has been widely used in the clinical treatment of lung cancer.As a key step in the process of treatment,the radiotherapy planning is the most time-consuming and least automated during the entire radiotherapy.In the process of designing the radiotherapy plan,the physician needs to manually delineate the organs at risk(OAR)and lesions.The efficiency and accuracy of the segmentation mainly depend on the clinical experience and medical knowledge of the physician and the physicist.The process of drawing is of heavy workload and little efficiency.Even more,the false delineation and missed diagnosis maybe appeared due to the human subjective factors.Traditional medical image processing algorithms require manual design features,and the processing steps are cumbersome,time-consuming,and less automated.In recent years,deep learning has been developed rapidly,and its powerful feature learning ability has made revolutionary progress in the fields of image processing,natural language processing,and speech recognition.Applying deep learning to medical image processing simplifies the complex process of traditional medical image processing and also has better performance in feature extraction.In order to improve the efficiency and quality of radiotherapy planning and lay the foundation for the automation design of radiotherapy plan,this paper introduces the problem of delineation for organs at risk(OARs)and potential lesion identification in the design of lung cancer radiotherapy planning,and applies deep learning to the planning for lung cancer radiotherapy.This paper uses deep learning to automatically segment OARs(the heart and lungs)for lung cancer radiotherapy planning,and to detect early lung cancer(the lung nodule).This paper first summarizes the development history and research status of deep learning,and introduces three application scenarios of deep learning in medical field: medical image segmentation,medical image recognition and computer-aided diagnosis.Then,aiming at the problem that procedure of OARs delineation in lung cancer radiotherapy plan is time-consuming and inefficient,the automatic parallel segmentation of OARs for lung cancer radiotherapy based on the dilated U-net neural network is proposed.A three-channel pseudo color image dataset are constructed with the lung window,heart window and mediastinum window,and it is divided into training set,verification set and test set.Then,the dilated U-net neural network is built,trained and optimized by training set and verification set.Finally,image segmentation performance is evaluated on the test set.The experimental results of lung segmentation shows that the average Dice similarity coefficient,average precision,average accuracy,and average recall of the dilated U-net network model are 97.48%,97.73%,99.1%,and 97.29%,respectively,and the results are slightly better than U-net network model;The experimental results of the heart segmentation results show that the average Dice similarity coefficient,average precision,average accuracy,and average recall of the dilated U-net network model are 92.11%,91.99%,94.54% and 92.49%,respectively,which are better than the U-net network model.The experimental results show that the method can effectively achieve the automatic parallel segmentation of lung and heart,and the segmentation result approximates the artificial delineation result.Compared with the U-net network model,the dilated U-net provides the similar performance for lung segmentation,and the superior segmentation ability for heart segmentation.Finally,for the early lung cancer,we investigate the lung nodule automatic detection based on the Faster R-CNN detection model.The LUNA16 dataset is used and processed as the target detection format dataset.To solve the problem that the lung nodule pixel area is very small which makes extracting its feature difficult,this paper resizes the input image,replaces the VGG middle net with the ResNet101 for extracting feature,and modifies the generation mechanism of the anchor according to small lung nodule.The experimental results show that expanding size of input image can improve the detection ability of detection model;the detection model with ResNet101 has the better ability for extracting feature,and it can identify the true lung nodule accurately;the small anchors after modifying have advantage to the generation of lung nodule proposals,and improve the comprehensive detection ability of lung nodule detection model.The average precision of modified lung nodule detection model is 88.91%.Therefore,this model can reduce the possibility of misdiagnosis of lung nodules,and quickly and efficiently detect lung nodules,which can help doctors to identify potential cancerous lesion during lung cancer radiotherapy planning.In summary,the proposed methods based on deep learning can achieve the automatic segmentation of OARs,and detect early stage lung cancer automatically for the planning for lung cancer radiotherapy.Therefore,the efficiency and quality of radiotherapy planning can be improved;the dependence of the plan quality on physician experience and the inefficiency of planning can be avoided.The above investigations may bring some foundations for the automation of lung cancer radiotherapy planning.