Author: Haijian Chen, Katja M. Langen, William Andrew LePain, Claire Tran, Mingyao Zhu 👨🔬
Affiliation: Emory Healthcare, Emory University, Georgia Institute of Technology 🌍
Purpose: To validate the performance of a commercial deep-learning segmentation (DLS) tool for head and neck cancer (HNC) and thoracic and abdominal cancer (TAC) by comparing it to manual segmentation for various organs at risk (OARs).
Methods: Anonymized data from 27 HNC patients and 20 TAC patients were used in a clinical treatment planning system (TPS). For HNC and TAC, 20 OARs and 11 OARs were evaluated respectively. DLS accuracy was quantitatively measured using the Dice score, which measured the volumetric overlap between the contours, and the Hausdorff distance, which calculated the maximum distance between two points from both contours.
Results: For HNC, DLS performed well for Brain, Brain Stem, Left and Right Eye, Mandibular Bone, Left and Right Parotid, and Left and Right Submandibular Glands, all with a mean Dice score 0.8 or higher and a mean Hausdorff distance less than 0.15 cm. For the TAC, the DLS performed well for Heart, Left and Right Lung, Left and Right Kidney, and the Liver, with an average Dice score 0.9 or higher and a mean Hausdorff distance less than 0.32 cm. For Cochlea and Trachea, our institutional standard differed from the TPS’s DLS training set and therefore had artificially decreased agreement. Similarly, for Spinal Cord and Esophagus, some manual contours only include the part close to the target and again have artificially decreased agreement.
Conclusion: This study revealed the effectiveness of a commercially available DLS for larger, geometrically simple OARs. Conversely, for smaller and anatomically complex OAR, the DLS model struggled to accurately predict the boundaries. In addition, differing segmentation methods between manual contours and DLS model for specific OARS like the esophagus and spina cord were also considered. The DLS is being implemented in routine clinical workflow for HNC and TAC patients.