3052 - Clinical Validation of Dose Regimen Generalized AI Prostate SBRT Plans: Dosimetric Comparison, Blinded Physician Review, and Deliverability Assessment

Purpose/Objective(s): To evaluate the clinical readiness of an AI-based automated planning system for prostate SBRT by comparing AI-generated plans against approved manual plans across dosimetric endpoints, blinded physician preference, and delivery quality assurance, while assessing generalization from the vendor-trained 36.25 Gy/5 regimen to the institutional 40 Gy/5 standard.

Materials/Methods: 23 prostate SBRT cases (40 Gy/5 fractions) were replanned with the AI system using identical CT and structure sets as approved clinical plans. Both plan sets were normalized to D95%=100%. The AI model was trained on 36.25 Gy/5 without institution-specific tuning. Target metrics included PTV D95%, D99%, and Paddick Conformity Index. OAR evaluation used an institutional SBRT constraint set; urethral structures were excluded from analysis given CT-only delineation and exclusion from model training. A blinded A/B preference test was conducted on 20 unique cases: urethral objectives were excluded from the evaluation scorecards, and plans were randomized as Plan A/B per-patient. Three physicians (a network attending, a faculty member, and a second-year resident) independently selected Prefer A, Prefer B, or No Preference. Five AI plans underwent patient-specific QA with gamma analysis at 2%/2mm criteria.

Results: AI plans met D95% in all cases with comparable conformality to manual plans. All AI plans used 2 arcs (vs 2–3 for manual), averaging 438 fewer MUs and <10-minute generation time. Excluding urethral structures, 99.2% of DVH objectives were met (390/393). In the blinded review (60 total evaluations), 42% were No Preference and 8% favored the AI plan (Table 1). When a preference was stated, manual plans were preferred 86%. Excluding urethra objectives, all AI plans were considered clinically acceptable by all reviewers. Dosimetrically, AI plans showed higher mid-dose rectal (D33% +149 cGy, D50% +131 cGy) and skin (+189 cGy) values on average, while meeting 99.2% of non-urethral constraints. Patient-specific QA demonstrated 100% gamma pass rates at 2%/2mm across all five plans tested (10/10 fields), with mean gamma values ranging from 0.15 to 0.24.

Conclusion: The AI-based automated planning system generated deliverable, clinically acceptable prostate SBRT plans with successful dose-regimen generalization and improved planning efficiency. Blinded physician review confirmed clinical equivalence for the majority of cases across non-urethral structures. Urethral sparing remains the primary limitation, attributable to CT-only delineation and exclusion from model training. These results support a path toward clinical implementation contingent on MR-informed urethral contouring and vendor optimization of urethral constraints.