Feasibility of automated AI-based contouring and stable radiomic feature assessment by HyperSight-CBCT Imaging for adaptive high-precision radiotherapy of prostate cancer

Abstract

Abstract This study evaluated segmentation accuracy, efficiency, and radiomic feature stability for manual (MD), artificial intelligence-based (AI), and hybrid (MD + AI) contouring of pelvic organs on planning CT (pCT) and HyperSight cone-beam CT (hCBCT) for adaptive radiotherapy. Dice similarity and 95th percentile Hausdorff distance (HD95) quantified segmentation agreement, while radiomic feature stability was assessed using the concordance correlation coefficient (CCC). Agreement between segmentation approaches was highest for bladder and femora (median Dice 0.95–0.96; HD95 1.88–2.17 mm), intermediate for prostate and rectum (median Dice 0.92; HD95 2.22–2.62 mm), and lowest for seminal vesicles and penile bulb (median Dice 0.76–0.83; HD95 3.01–3.41 mm). AI and MD + AI reduced contouring times by about 90% and 60% compared to MD. Radiomic feature stability differed significantly between segmentation modes (all p adj ≤ 0.05). GLRLM features exhibited significantly higher stability than other features, whereas morphological features showed lower stability. Median radiomic feature stability was highest for bladder and femora, and intermediate for prostate and rectum. In conclusion, AI-based and hybrid contouring achieved high accuracy and substantial time savings, while texture- and intensity-based radiomic features showed robustness with AI segmentation. This study demonstrated feasibility of extracting distinct, reliable quantitative parameters based on AI-only contouring of pelvic structures.