Development And Validation Of A Deep Learning Algorithm For Auto-Delineation Of Clinical Target Volume And Organs At Risk In Cervical Cancer Radiotherapy

Abstract

The delineation of clinical target volume (CTV) and organ at risks (OARs) is an inevitable, laborious and subjective step in cervical cancer radiotherapy. The aim of this study was to propose and evaluate a novel end-to-end convolutional neural network (CNN) for fully automatic and accurate cervical cancer CTV and OARs delineation. CT images of 104 locally advanced cervical cancer patients were selected. We developed a novel CNN network, called radiotherapy delineation network (RTD-Net). It is built upon the popular U-network and extended in two different ways: 1) The dual path network (DPN) has been introduced into the U-Net to promote the feature learning ability and make the segmentation more accurate. 2) A self-adaptive weighted (SAW)loss function has been designed to tackle the different levels of class-imbalance problem. We performed comprehensive comparisons and experiments between the predicted and manually delineated contours to assess our method. The dice similarity coefficient (DSC) of RTD-Net was 86.2±7.7% for bladder, 82.9±7.5% for bone marrow, 79.1±9.5% for left femoral head, 83.1±5.8% for right femoral head, 67.9±6.2% for rectum, 82.8±4.2% for small intestine, 79.0±7.5% for spinal cord and 84.1±3.9% for CTV, respectively. We also conducted clinical expert’s subjective assessment experiment and 87.2% predicted contours require no further revision. The average delineation time for one patient CT image was within 15 seconds. The CTV and OARs delineated by RTD-Net provided close agreement to the ground truth compared with current inter-observer variability. RTD-Net could significantly reduce the contouring time.