Evaluation of general and specialized knowledge-based planning models for prostate cancer with bowel proximity.

This study evaluated knowledge-based planning (KBP) models for prostate cancer, focusing on cases where the bowel lies close to the planning target volume excluding the rectum (PTV-R). Such proximity increases toxicity risk, raising the question of whether a Special Model (Model-S) is required. Although General Models are widely applicable, their performance in bowel proximity cases remains uncertain. A General Model (Model-G) was trained with 30 standard cases and a Model-S with 30 bowel proximity cases. Thirteen additional proximity cases were used for independent evaluation. RapidPlan-generated plans (Plan-G and Plan-S) were compared with clinically used manual plans (Plan-M). Plan quality was evaluated using D98% and D2% for targets, dose-volume indices for the rectum and bladder, and D0.1/D1 cc for the bowel. Statistical analyses employed the Mann-Whitney U and paired t-tests (p < 0.05). In the independent evaluation, Plan-M achieved significantly lower D2% than Plan-G (p < 0.001), while Plan-S showed no significant difference (P = 0.069). Organ-at-risk sparing was comparable across all plans (p > 0.05). Among the 13 evaluation cases, one Plan-G slightly exceeded the bowel D1 cc constraint of 60 Gy, indicating that in such bowel-descended cases, manual re-optimization or use of the Special Model may be required, defining the boundary where General Models remain reliable. Both models yielded clinically acceptable plans, and no clear dosimetric advantage was observed for the Special Model. These findings demonstrate the robustness of KBP against anatomical variability and highlight its potential as a foundation for future AI-assisted adaptive radiotherapy frameworks.