TU-AB-303-07: Evaluation of Automatic Segmentation of Critical Structures for Head-And-Neck and Thoracic Radiotherapy Planning

Abstract

Purpose: To evaluate the performance of commercially available automatic segmentation tools built into treatment planning systems (TPS) in terms of their segmentation accuracy and flexibility in customization. Methods: Twelve head-and-neck cancer patients and twelve thoracic cancer patients were retrospectively selected to benchmark the model-based segmentation (MBS) and atlas-based segmentation (ABS) in RayStation TPS and the Smart Probabilistic Image Contouring Engine (SPICE) in Pinnacle TPS. Multi-atlas contouring service (MACS) that was developed in-house as a plug-in of Pinnacle TPS was evaluated as well. Manual contours used in clinic were reviewed and modified for consistency and served as ground truth for the evaluation. Head-and-neck evaluation included six regions of interest (ROIs): left and right parotid glands, brainstem, spinal cord, mandible, and submandibular glands. Thoracic evaluation includes seven ROIs: left and right lungs, spinal cord, heart, esophagus, and left and right brachial plexus. Auto-segmented contours were compared with the manual contours using the Dice similarity coefficient (DSC) and the mean surface distance (MSD). Results: In head- and-neck evaluation, only mandible has a high accuracy in all segmentations (DSC>85%); SPICE achieved DSC>70% for parotid glands; MACS achieved this for both parotid glands and submandibular glands; and RayStation ABS achieved this for spinal cord. In thoracic evaluation, SPICE achieved the best in lung and heart segmentation, while MACS achieved the best for all other structures. The less distinguishable structures on CT images, such as brainstem, spinal cord, parotid glands, submandibular glands, esophagus, and brachial plexus, showed great variability in different segmentation tools (mostly DSC 3mm). The template for RayStation ABS can be easily customized by users, while RayStation MBS and SPICE rely on the vendors to provide the templates/models. Conclusion: Great variability was observed in different segmentation tools applied to different structures. These commercially-available segmentation tools should be carefully evaluated before clinical use.