Time efficiency, geometric accuracy, and clinical impact of AI-assisted contouring of organs at risk in head and neck cancer radiotherapy.

Abstract

Background and purpose: Ensuring the reliability and accuracy of artificial intelligence (AI)-generated contours is paramount, as discrepancies could lead to inadequate protection of healthy tissues. With increasing clinical workload, the aim of this study was to assess the time-saving potential of AI-assisted organs at risk (OAR) contouring in head and neck cancer (HNC) treatment planning, while also evaluating geometric accuracy, variability, and dosimetric impact. Patient/material and methods: Twenty patients had 12 OAR contoured by 11 certified dosimetrists and ARTplan (Therapanacea), including the brainstem, cochleas, larynx, mandible, oral cavity, parotid glands, pharynx constrictor muscles, spinal cord, right submandibular gland and thyroid gland. Comparisons were made using geometrical metrics, including Mean Surface Distance, Dice Similarity Coefficient (DSC), Hausdorff Distance, Volume Difference, and Centre of Mass Difference, as well as relevant dose-volume metrics, and total contouring time.

Results: Median manual contouring time of the OARs was 55 (range: 17-151) minutes per patient, while adjusted AI-based structures required 17 (7-42), resulting in 69% time saved. For manual, adjusted and AI-contours, the mean DSC were generally high, averaging 0.85, 0.86, and 0.81 respectively across the evaluated structures. Notably, variability was lowest for the AI and adjusted contours. Average mean and max dose differences were acceptably low (<3.2 Gy) for all OARs.

Interpretation: The results support the integration of AI-based contouring in HNC treatment planning. With minor adjustments, the contours achieve very good clinical quality and demonstrate improved consistency compared to manual contours, while significantly reducing contouring time.