Evaluation of Deep Learning-Based Auto-Segmentation of Target Volume and Organs-at-Risk in Breast Cancer Patients

Abstract

In breast cancer patients receiving radiotherapy (RT), lower radiation doses to the nearby normal organs can reduce treatment-related toxicities. Also, the importance of meticulous target delineation is increasing due to the advancement of RT modalities such as intensity-modulated RT. Subsequently, accurate target volume and organs-at-risk (OAR) segmentation has become imperative, adding considerable amount of workload in the clinical field. Thus, deep learning-based auto-segmentation can be an expedient tool for target and OAR segmentation. Here, we evaluated the deep learning-based auto-segmented contours compared to manually delineated contours in breast cancer patients. Clinical target volumes (CTV) for bilateral breasts and regional lymph node and OARs including heart, lung, esophagus, spinal cord and thyroid were manually delineated on a planning computed tomography scans of 61 breast cancer patients who received breast-conserving surgery. In addition, cardiac substructures including right coronary artery (RCA), left anterior descending artery (LAD), atriums and ventricles were manually delineated. Afterwards, a two-stage convolutional neural network (CNN) algorithm was conducted to effectively segment the organs. Quantitative metrics, including dice similarity coefficient (DSC) and Hausdorff distance (HD), and qualitative scoring by expert and non-expert panel were used for analysis. Inter-observer variability was also assessed for contours by three radiation oncologists on a randomly selected case. The correlation between the auto-segmented and manual contours was good for OARs including heart, lung, esophagus, spinal cord, thyroid and bilateral atriums and ventricles with mean DSC higher than 0.80. In addition, CTVs also showed favorable results with mean DSC higher than 0.70 for all breast and regional lymph node CTVs. However, auto-segmented contours of RCA and LAD showed reduced performance with mean DSC lower than 0.5 and mean HD higher than 20 mm. Qualitative subjective scoring by physicians showed good results for all CTV and OARs. As for inter-observer variability, only heart showed DSC of 0.91, whereas the other OARs showed DSCs lower than 0.80. For CTV, although breast CTV showed an acceptable mean DSC of 0.85, other CTVs for regional lymph nodes showed poor results with mean DSC ranging from 0.45 to 0.75. The feasibility of deep learning-based auto-segmentation was shown in this study. Although deep learning-based auto-segmentation cannot be a substitution for radiation oncologists, it can be an expedient tool in clinic, by assisting radiation oncologists and consequently enhancing the quality control of radiotherapy.