Comparative analysis of AI-generated and deformed image registration contours on daily CBCT in prostate cancer...

Deep learning (DL)-based auto-segmentation has rapidly become the state-of-the-art in radiotherapy planning, significantly reducing contouring time while achieving geometric accuracy comparable to expert-derived contours [1-3]. While AI contouring on CTp is now widely established, its application to cone-beam CT (CBCT) is less well explored, despite CBCT's critical role in daily image guidance for prostate radiotherapy. Current adaptive workflows rely on manual contouring or deformable image registration (DIR), both of which are resource-intensive and subject to limitations in accuracy and consistency. Recent advances in AI-based CBCT segmentation have shown promise in reducing manual workload, improving contour consistency, and supporting adaptive radiotherapy (ART) workflows [4]. To assess the clinical implications of these developments, this study retrospectively analyzed CBCT images from 20 prostate cancer patients, comparing AI- and DIR-generated contours to evaluate systematic differences and their potential impact on dosimetry and ART decision-making.

Twenty prostate radiotherapy patients were retrospectively selected, treated with either 42.7 Gy in 7 fractions or 60 Gy in 20 fractions, and imaged on Halcyon linear accelerators using Hypersight CBCT ([Formula: see text]). AI-generated contours were produced with Limbus AI v1.8.0, while deformable image registration (DIR) contours were propagated from planning CTs in Velocity v4.2. Contour accuracy was assessed by two senior medical officers using a four-point Likert scale across 140 CBCTs. Prostate, bladder, and rectum were analyzed using Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), mean surface distance (MSD), center-of-mass (COM) displacement, and volumetric change relative to the planning CT. Dosimetric evaluation included [Formula: see text], [Formula: see text], [Formula: see text], and clinically defined organ-at-risk metrics to assess potential implications for adaptive radiotherapy. Statistical significance was tested using paired Student's t-tests and Wilcoxon signed-rank tests with a threshold of [Formula: see text].

AI-generated contours achieved acceptable clinical accuracy in >80% of cases, with fewer severe or medium errors compared to DIR-derived contours, which required minimal changes of 49%. Quantitative analysis demonstrated broadly comparable Dice Similarity Coefficients (DSC), Hausdorff Distance (HD), and mean surface distance (MSD) across prostate, bladder, and rectum. Organ variation on CBCT revealed larger mean centre of mass shifts and volume differences for AI, particularly in bladder contours, whereas DIR showed smaller systematic deviations. Dosimetric comparisons highlighted that prostate dose metrics were significantly different between methods, while bladder differences were mostly non-significant except at high-dose volumes, and rectum analysis revealed consistent statistically significant variations. Overall, although both methods captured daily anatomical changes, suggesting complementary strengths depending on adaptive radiotherapy application.

AI-generated contours for prostate radiotherapy on CBCT images demonstrate high geometric accuracy and clinical usability, requiring minimal expert correction, while DIR contours, although generally usable, show greater variability, particularly for organs subject to large anatomical changes such as the bladder and rectum. Despite similar geometric comparisons, statistically significant dosimetric differences highlight the importance of careful expert verification, especially for sensitive structures like the rectum. These findings support the integration of AI-based contouring into adaptive radiotherapy workflows to streamline clinical processes, reduce workload, and maintain treatment accuracy, while emphasizing that automated contours, whether AI- or DIR-derived, should always undergo expert review to ensure safe and effective patient care.