Comparative Study And Application Of Deep Learning-based Auto-segmentation In Cervical Cancer Radiotherapy

Cervical carcinoma has higher incidence rate for women in the world and radiother-apy by X-rays is commonly served.Clinicians need to spend a lot of time to complete the delineation of targets and organs at risk during radiotherapy,therefore,a number of tech-niques focusing on auto-segmentationhave been developed by experts in the field of ra-diotherapy.Although some commercial systems have implemented auto-segmentation technology based on deep learning or prior knowledge,there are still crucial issues un-der exploration,such as,is the deep learning-based method fully ahead when compared with traditional auto-segmentation technology and what are the advantages in clinical applications?The aim of this doctoral thesis is to explore the advantage and usability of the auto-segmentation technique based on deep learning in the clinical treatment of cervical cancer by radiotherapy.The specific research goal is to comprehensively evaluate the current mainstream auto-segmentation technology and demonstrate its clinical applica-tion effect.To achieve the above goals,three specific tasks have been addressed:(1)To analyze the deep learning-based technology in a systematic perspective and compare the distinctions with knowledge-based technology in literature;(2)To compare sep-arately the accuracy differences for the manual delineation performances from junior doctors and the deep learning-based auto-segmentation technology with a gold stan-dard of manual delineation of chief physician combined with the current clinical status in China;(3)Deep learning-based auto-segmentation model for cervical cancer targets trained by local data is integrated with the commercial contour system Deep Viewer.The results show that:(1)The previous traditional auto-segmentation methods can facilitate the segmentation of OAR in some extent,it is necessary to establish a large number of atlases covering specific human parts.The quantity of atlases and the quality of contour will have a critical effect on the segmentation accuracy,which is a huge lim-itation when applying on clinical work;Meanwhile,prior knowledge-based technology lags behind deep learning-based technology in segmentation time,objective evaluation(DSC,centroid deviation,HD₉5,average surface distance)and subjective evaluation of clinicians(Turing test and grade evaluation of segmentation result),especially for or-gans such as rectum segmented by prior knowledge-based technology,the segmentation ability based on deep learning has been improved a lot.(2)By comparing and analyzing the differences between auto-segmentation technology based on improved U-Net con-volution neural network model and the results of delineation of targets and OARs from the junior doctors,we find that auto-segmentation technology is superior to junior doc-tors in DSC of CTV,femoral head,and small intestine,among which auto-segmentation DSC of CTV is(0.861±0.020),while that of junior doctors is(0.829±0.017).It shows that the segmentation ability of the deep learning-based auto-segmentation model was not worse than that of junior doctors,which proved the clinical applicability of the im-proved U-Net convolution neural network model adopted in this study.(3)This project successfully integrated the model into the commercial contour system Deep Viewer,and realized the auto-segmentation of cervical cancer targets.