Artificial intelligence based auto-contouring solutions for use in radiotherapy treatment planning of head and neck cancer

Abstract

BackgroundManual contouring is time-consuming and subjective. Thus, auto-segmentation methods, which can be deployed in the existing workflow, are needed. The objective of this study was to assess the feasibility of Limbus AI and AI Rad Companion auto-contours for head and neck treatment planning. MethodsHead and neck patients treated with RapidArc were selected retrospectively. The manual contours on the planning CT were used as reference. Geometric analysis of the auto-contours was performed using several evaluation metrics such as the Dice Similarity Coefficient (DSC) and the Mean Distance to Conformity (MDC). Dosimetric analysis was performed by recalculating the original plan on the auto-contours and comparing Dose Volume Histogram (DVH) metrics to the original plan. Results and discussionBoth AI tools tend to underestimate the volumes of brainstem and cord. For brainstem and parotids, median DSC values were ≥ 0.8. For all auto-contours, median MDC values were ∼ 3–6 mm. Median differences were found of up to ±7 % in DVH points on the auto-contours relative to the planning CT contours, but these were not statistically-significant. ConclusionThe auto-contours could be used as a starting point to assist the clinician with the manual contouring of structures on the planning and re-scanning planning CT.