S2018
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
originally trained on 36.25 Gy/5. Target metrics included PTV D95%, D99%, and the Paddick Conformity Index. Organ-at-risk (OAR) evaluation used the institutional SBRT constraint set, including urethra D0.035 cc and urethra+3 mm, bladder wall, rectal wall, femoral heads, penile bulb, and skin. Plan quality was summarized as deltas from constraint thresholds. Urethra contours were CT-based; MR was not available, and urethra/urethra+3 mm were not part of the vendor's model training targets. Results AI plans achieved robust target coverage (D95% met in all cases; mean D99% was higher by 317 cGy) with conformality index comparable to Manual. The AI generalized to 40 Gy without retuning, preserving target coverage and producing favorable trade-offs for some OARs (mean femoral head sparing improved by ~179 cGy; modest reductions in bladder/rectal maxima). All AI plans consistently used 2 arcs compared to Human plans which varied between 2 and 3 arcs, with AI plans averaging 438 fewer monitor units and requiring less than 10 minutes for plan generation. Constraint deviations were concentrated in structures not included in model training: all AI plans exceeded the urethral D0.035 cc limit at 40 Gy by a mean of ~101 cGy, and urethra+3 mm/penile bulb contributed additional violations (93.5% of the total DVH objectives passed across 23 AI plans). AI plans also showed increased dose to skin and mid-to-low rectal dose regions.
Conclusion: Automated planning produced rapid, single-fraction palliative bone plans with geometric accuracy, dosimetric quality, and QA performance comparable or superior to ClinicalPlans, while reducing plan complexity. Lower Dice values in ribs and at C6–D2 likely reflect greater anatomical curvature and variability at the neck–shoulder transition, where clinical and automatic CTV definitions diverge more. Overall, the data substantiate a short, ~1-hour CT-to- RT workflow for palliative indications, improving access and potentially accelerating symptom relief without compromising plan quality or safety. Keywords: AutomaticPlanning Clinical validation of AI prostate SBRT plans: comparison to manual plans and generalization from 36.25 Gy to 40 Gy Nayoon J Lee 1,2 , Anran Zhao 1 , Ledi Wang 1 , Steven Philbrook 1 , William R Green 1 , Madalina-Liana Costea 3 , Baris Ungun 4 , Remi Vauclin 4 , Elie Mengin 4 , Nikos Paragyos 5 , Rafe A McBeth 1 1 Radiation Oncology, Perelman School of Medicine, Philadelphia, USA. 2 Physics and Astronomy, University of Pennsylvania, Philadelphia, USA. 3 Clinical Affairs, TheraPanacea, Paris, France. 4 Physics, TheraPanacea, Paris, France. 5 Chief Executive Officer, TheraPanacea, Paris, France Purpose/Objective To assess institutional clinical readiness of TheraPanacea SmartPlan for prostate SBRT by comparing AI-generated plans with approved manual plans and evaluating generalization from the vendor- trained 36.25 Gy/5 regimen to the local 40 Gy/5 standard. Material/Methods Digital Poster 5006 23 prostate SBRT cases were replanned with SmartPlan ("AI") using the same CT and clinical structure sets as the approved plans ("Human"). Prescriptions were 40 Gy in 5 fractions, with both Human and AI plans normalized to 100% dose covering 95% of the target volume; the AI model was
Conclusion SmartPlan produced clinically competitive target coverage with successful generalization from 36.25 Gy/5 to 40 Gy/5, demonstrating improved femoral head sparing and planning efficiency (438 fewer MUs, <10-minute generation time). Urethral sparing represented the primary limitation, anticipated given CT-only delineation and exclusion of urethra structures from model training. Clinical implementation is feasible with two targeted
Made with FlippingBook - Share PDF online