Target Re-Delineation in Daily Online Adaptive Radiotherapy of Bladder Cancer: AI vs. Adapters

Abstract

The purpose was to assess the quality of artificial intelligence (AI) generated targets in cone-beam CT (CBCT) guided online adaptive radiotherapy (oART) of bladder cancer and compare it to corresponding contours edited by adapters (radiotherapy technicians performing oART routinely). Furthermore, to assess the inter-observer delineation variation in daily target re-delineation in oART of bladder cancer. Ten consecutive patients treated with oART for muscle-invasive bladder cancer from Feb 2020 to Feb 2021 were retrospectively included. One CBCT randomly selected from each of the 10 clinical treatment courses was exported to an emulator, where delineations and oART simulations were carried out. The bladder was automatically segmented (clinical target volume (CTV)) using AI. Seven adapters independently reviewed and, if necessary, manually edited the AI-generated structure. A ground truth structure was first delineated blindly by a senior oncologist, then reviewed by an oART experienced medical physicist, and finally edited to a consensus structure. Planning CT and MR images with reference delineations were available during all simulations and delineations. The AI-generated and adapter-edited CTVs were compared to ground truth using dice similarity coefficient (DSC) and volume difference. The inter-observer variation in adapter-edited CTVs was assessed using coefficient of variation (CV) and the generalized conformity index (CIgen). Cigen was calculated as the ratio of the sum of all overlapping volumes between pairs of observers and the sum of all overlapping and non-overlapping volumes between the same pairs. Nine CTVs (7 adapter-edited, 1 AI-generated, 1 ground truth) were generated per patient. Patient cases included three male and seven females; three had catheter. The ground truth volume ranged from 51.5 to 221.6 cc. The median volume difference compared to ground truth was -4.5 (range, -17.8 - 42.4) cc and -15.5 (range, -54.2 - 4.3) cc for adapter-edited and AI-generated structures, respectively. Corresponding DSC values were 0.87 (range, 0.71 - 0.95) and 0.84 (range, 0.64 - 0.95). The AI-generated CTV was smaller than ground truth for all patients except one and smaller than all adapter-edited CTVs. The largest differences among adapter-edited CTVs were observed in cranial, caudal and posterior directions. The median CV was 0.08 (range, 0.05-0.11) and Cigen was 0.78 (range, 0.71-0.88). Target re-delineations in daily CBCT-based oART of bladder cancer demonstrated small inter-observer variation. Manual adjustment of AI-generated structures resulted in improved accuracy in target delineation compared to ground truth. Efforts to reduce the inter-observer variation, e.g., physician support during the initial fractions and review by a second adapter, have been implemented clinically.