(OA40) Deep Learning-Based Autosegmentation for Head and Neck Radiotherapy

Abstract

To investigate the feasibility of implementing a deep learning-based autosegmentation system for head and neck cancer patients undergoing radiotherapy. The accuracy of four different contouring approaches was investigated. Two atlas-based autosegmentation (ABAS) systems, MIM (MIM Software, Inc., Cleveland, OH) (ABAS1) and RayStation (RaySearch Laboratories AB, Stockholm, Sweden) (ABAS2), one deep learning-based autosegmentation (DLBAS) system Nimble ContourTM (Nimble Therapy, LLC, San Francisco, CA), and intra-physician contouring variability (IPCV), were evaluated. 200 patients were considered in this study. 160 patients were used to train the ABAS1 and DLBAS algorithms; due to limitations of the system only 10 patients were used to train ABAS2. The IPCV was generated by expert manual contouring of the spinal cord, mandible, right parotid gland, left parotid gland, oral cavity, brainstem, larynx, esophagus, right submandibular gland, left submandibular gland, right temporomandibular joint (TMJ), and left TMJ for five head and neck cancer patients ten times each. Contouring approaches were analyzed by calculating the average Dice coefficients, root mean squared error (RMSE), maximum dose difference, and mean dose difference. The IPCV scored the best for most evaluation metrics. DLBAS performed only slightly worse than IPCV for most evaluation metrics and was able to match or beat the IPCV for some OARs. DLBAS was significantly better than ABAS1 and ABAS2 for most evaluation metrics. The ABAS systems required approximately 1.5 minutes per training patient; the DLBAS method required approximately 72 hours to train all 12 OARs. By moving towards deep learning-based algorithms, autosegmentation accuracy can approach ICPV accuracy. Deep learning-based autosegmentation is a superior alternative to atlas-based autosegmentation, given sufficient training data and training time. This technique can help reduce physician workload and standardize practice between institutions.