Abstract
PURPOSE: In this prospective cross-over study, the precision of manual correction of the clinical target volume (CTV) during online-adaptive radiotherapy (ART) of patients with bladder cancer was investigated in dependence on the input of applied computational algorithms. METHODS AND MATERIALS: Online-adaptive-radiation therapy (ART) fractions were analysed from a prospective registry as a component of a trimodality treatment for bladder cancer. The CBCT-guided adaptive radiation therapy (CBCTgART) workflow uses a CBCT acquired at the beginning of a treatment session (CBCT1) to update the target volume contours of the planning CT (pCT) on the current anatomy of the day. The study comprises radiotherapy fractions with target (CTV) segments showing clinically relevant deviations (>5 mm) when mapped between pre-treatment (CBCT1) and post-adaptation (CBCT2) images. CTV(algo1-2) were obtained from two deformable-image-registration algorithms (algo(1,) algo(2)) based on AI-generated normal-tissue contours and a rigid copy of the CTV(original) termed CTV(algo3) (algo(3)). Verification cone-beam-computed-tomography (CBCT2) acquired during ART showed CTV(algo1-3) on new anatomical scenarios. The task of 13 physicians was to adjust CTV(algo1-3) on CBCT2 to match the CTV(original) on CBCT1. RESULTS: In 151/169 pairwise comparisons, mean-distance-of-segments-to-ground-truth (MDA(GT)) after adjustment by a physician was larger for the poorer algorithms than for the better ones. The binomial proportion was 0.893, significantly different from 0.5 as null proportion (p < 0.0001, exact-binomial-test), and not dependent on the physician or on the respective patient. In a quantitative analysis by a mixed-model, the MDA(GT) of the manually adjusted segments between 16% and 50% of the errors were on average corrected by the physicians, with interindividual differences (p < 0.0001,F-test). CONCLUSIONS: The present study results reveal that the quality of AI-generated contours is of crucial importance for a proper online onboard ART, as errors may bias clinical workflow and contouring. The accuracy of supervised structures strongly depends on the accuracy of the AI generated contours before supervision.