S1827
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
ML and reference (DMPO) plans by physicists and physicians;Dosimetric comparison usingDVHs, dose- volume metrics, clinical goals, index (homogeneity, conformity, PQI), and robustness to deformations;Plan deliverability and complexity verification using ArcCheck ( ≥ 95% passing rate, 2%/2 mm) and modulation complexity score /MU analysis. A script was developed to automate the planning workflow: structure creation, deformed CTs generation for robust scenarios, and ML-based optimisation. Final ML plans were reviewed and approved by physicians and physicists prior to clinical implementation. Results: ML-generated plans demonstrated excellent dosimetric agreement with reference plans for PTV coverage (V47.5Gy>97.31% vs 97.70%).OAR doses (lung, heart, A_LAD) were lower in ML plans, although some showed slightly higher hotspots (up to 107%), requiring one post-processing run.
deviation from the optimal dose distribution. High dose maxima at level 4 resulted in skin doses beyond the known dose guidances. While this pilot study establishes technical feasibility of SIB dose escalation in ultrahypofractionated naRT for extremity STS, prospective clinical studies are needed to investigate clinical safety. Keywords: boost dose escalation, sarcoma, hypofractionation Digital Poster Highlight 1304 Implementation and clinical validation of a machine learning–based planning model for left breast intensity modulated radiation therapy treatments. Célia Petitjean 1,2 , Jocelyne Mazurier 1 , Xavier Franceries 2 , Susanna Adera 1 , Jéremy Bétend 1 , Jeremy Camilleri 1 , Vincent Connord 1 , Nicolas Mathy 1 , Fanny Solinhac 1 , Olivier Gallocher 1 , Guillaume Janoray 1 , Gaëlle Jimenez 1 , Igor Latorzeff 1 , Baptiste Pichon 1 , Baptiste Pinel 1 1 Orion - service de radiothérapie, Clinique Pasteur, Toulouse, France. 2 ToNIC Toulouse NeuroImaging Center, Université de Toulouse, INSERM, Toulouse, France Purpose/Objective: This study aims to present a methodology for implementing machine learning (ML) models in clinical practice, applied to the automation of treatment plan generation for left breast IMRT cases. Material/Methods: RaySearch provides generic ML-based planning models that can be adapted to local clinical settings by adjusting prediction (dose distribution) and mimicking parameters (delivery settings) during a commissioning phase.These ML models have been integrated into our radiotherapy service since 2023 for prostate, benefiting 408 patients and achieving plan quality similar to DMPO-optimisation, while reducing planning time by a factor 3 (ML+DMPO post-processing) to 7 (ML only). Building on this experience, the approach was extended to left breast treatment planning, which represents a major workload in routine practice.This study focuses on commissioning and adjusting a ML model for automatic generation of IMRT treatment plans for left breasts without lymph node involvement, prescribed 50Gy+16Gy (boost) on Halcyon, using DIBH. A robust planning strategy was incorporated to account for anatomical variations (breast swelling/deflation), with robust objectives for PTV applied on the reference CT.The commissioning used a local database of 23 patients and followed 4 steps:Local parameters adjustment to reproduce reference plans quality,Blind expert review comparing
Robustness analysis demonstrated plan stability similar to reference plans. All ML plans met ArcCheck deliverability criteria and exhibited similar complexity metrics.The complete workflow was generated and optimized in 23 minutes on average (passive time), saving time by a factor 4. Conclusion: This study confirms the feasibility and reliability of ML- based automatic planning for left breast IMRT treatments, providing plan quality comparable to conventional optimization with significant time savings. Ongoing work aims to extend this ML-based workflow to include cases with lymph node irradiation. References: [1]J.-H. Bolten et al., ‘A fully automated machine- learning-based workflow for radiation treatment planning in prostate cancer’, ctRO, vol. 52, May 2025, doi: 10.1016/j.ctro.2025.100933.[2]M. Zeverino et al., ‘Clinical implementation of deep learning-based
Made with FlippingBook - Share PDF online