S1937
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
planning remains limited compared to auto- contouring. Plan Feasibility (Sun Nuclear Corporation, Melbourne, FL, USA) predicts the ideal achievable dose distribution for each patient geometry, by ray-tracing outward from the target edge and applying measured leaf-edge fall-off kernels, creating a physically realistic dose gradient. It defines how good a plan could theoretically be, but not how to reach it. PlanAI (formerly Oncospace, now part of SNC) addresses this gap. Trained on over 5,000 clinically delivered plans from Johns Hopkins University, it uses a gradient- boosting ensemble model to predict the best achievable dose-volume histograms (DVHs) and generate corresponding optimisation objectives that can be imported directly into RayStation or Eclipse. Together, Plan Feasibility defines what is possible; PlanAI proposes how to achieve it. This study reports the first international evaluation of PlanAI, assessing whether AI-generated objectives can improve or standardise plan quality, and whether these improvements align with Plan Feasibility predictions. Material/Methods: Thirty previously treated cases (15 prostate and 15 lung) were retrospectively analysed. Baseline clinical plans were exported with associated Plan Feasibility metrics. Each case was processed through PlanAI via a RayStation script (CT + RTSS upload, template selection, import of AI-generated objectives). A single, unedited optimisation run (no additional human optimisation) was executed for each plan, reflecting AI performance. Dose-volume metrics (PTV V95, Dmax, OAR Vx and Dcc) were compared with human- optimised plans. Results: For the first five prostate cases, PlanAI maintained clinically acceptable PTV coverage ( Δ V95 = − 3.6 to +0.3 pp, median − 0.5 pp). Across 65 OAR constraints (13 constraints × 5 cases), PlanAI delivered lower or equal dose in 77 % of comparisons, confirming consistent OAR sparing. As illustrated in Figure 1, rectal and bladder V18 and V20 Gy volumes were typically 1-5% lower; while femoral head D10cc decreased by 1-3 Gy. One bowel-adjacent outlier exhibited higher low-dose spread.
An example DVH is shown in Figure 2, with solid lines PlanAI and dashed lines human-optimised plans. All AI plans were generated in a single optimisation cycle and further human refinement would likely enhance these results.
Conclusion: PlanAI achieved non-inferior target coverage and consistent OAR sparing after a single, unattended optimisation, demonstrating its potential to standardise and accelerate planning across institutions. These early findings position AI-assisted objective generation as a practical bridge to full scale automated planning. References: [1] Doolan P J, Charalambous S, Roussakis Y, Leczynski A, Peratikou M, Benjamin M, Ferentinos K, Strouthos I, Zamboglou C and Karagiannis E 2023 A clinical evaluation of the performance of five commercial artificial intelligence contouring systems for radiotherapy Front. Oncol.13 1–13[2] Hurkmans C, Bibault J E, Brock K K, van Elmpt W, Feng M, David Fuller C, Jereczek-Fossa B A, Korreman S, Landry G, Madesta F, Mayo C, McWilliam A, Moura F, Muren L P, El Naqa I, Seuntjens J, Valentini V and Velec M 2024 A joint ESTRO and AAPM guideline for development, clinical validation and reporting of artificial intelligence models in radiation therapy Radiother. Oncol.197 110345 Keywords: AI, treatment planning, evaluation
Made with FlippingBook - Share PDF online