ESTRO 2025 - Abstract Book

S3421

Physics - Machine learning models and clinical applications

ESTRO 2025

Conclusion: Our study demonstrated that physician-selection of ML-generated RT plans for treatment was high and that ML generated RT plans maintain low toxicity levels with no significant GU Grade 2+ differences compared with human generated RT plans. Results suggest ML planning with an option for human-generated RT plans maintains favorable clinical outcomes and should encourage further adoption as a standard-of-care.

Keywords: Machine Learning, Treatment Planning, Outcomes

References: 1 McIntosh C, Conroy L et al. Clinical integration of machine learning for curative-intent radiation treatment of patients with prostate cancer. Nat Med. 2021 Jun;27(6):999-1005.

3355

Digital Poster Development and Validation of a Knowledge-Based Model for Automated CyberKnife Planning in Brain Lesions Rita C. Buono, Roberta Castriconi, Claudio Fiorino, Alessia Tudda, Carmen Gigliotti, Lucia Perna, Antonella Del Vecchio, Sara Broggi Medical Physics, IRCCS San Raffaele Hospital, Milan, Italy Purpose/Objective: The Knowledge-Based (KB) approach in automated planning systematically organizes and stores essential information to guide treatment plan optimization. The current study aimed to evaluate the correlation between the equivalent radius of the PTV and the Body volume receiving specific doses (isodose shells) and to translate this information into a robust KB methodology for automatic plan optimization with CyberKnife (CK) for single brain lesions. Material/Methods: We retrospectively analyzed 60 plans for single lesions not adjacent to OARs (excepting the brain) treated at our Institute in 1-5 fractions (14-40Gy). Dose Gradient Index (DGI) and Conformity Index (CI) were calculated for each plan. Plans with CI<1.2 were selected to create the model. A linear regression analysis assessed the relationship between the equivalent radius of the PTV [1] and the Body volume receiving pre-defined dose values (100%, 85%, 65%, 50%, 40%, and 30% of the prescribed dose), chosen to accurately describe the dose fall-off outside PTV. This fitting was used to estimate isodose shell radii based on PTV volume for new patients; an optimization template was then fine- tuned to obtain optimal solutions, mimicking “best” planners, based on the 80% highest confidence interval of the predicted DGI values. The KB model was tested on 13 new random patients, comparing automatic plans (without additional planner intervention) against previously delivered clinical plans. Results: The KB model showed lower or similar CI values compared to clinical (1,11 ± 0,05 vs 1,25 ±0,16, p-value= 0.0059 Wilcoxon test), reflecting better or equivalent target conformity (Figure 1, Right). It also achieved higher DGI values (88±15% vs 79±17%, p-value=0.00012 Wilcoxon test), Figure 1, Left. The KB model significantly reduced radiation exposure to healthy brain tissue (p<0.05), particularly at lower isodose levels, as shown by the spared Brain-CTV volume compared to clinical plans (Figure 2). This enhanced brain sparing was achieved without compromising target coverage. Additionally, CK machine parameters suggested that KB-generated plans were generally more efficient in terms of number of beams (66±30 vs 100±49, p-value= 0.01746 Wilcoxon test) and delivery times (32±9 vs 36±11, p-value=0.2481 Wilcoxon test).