S1971
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
References: 1. Zeverino M, Fabiano S, Jeanneret-Sozzi W, Bourhis J, Bochud F, Moeckli R. Enhancing automated right-sided early-stage breast cancer treatments via deep learning model adaptation without additional training. Med Phys. 2025 May;52(5):3280-3297. doi: 10.1002/mp.17682 Keywords: Auto-planning,Deep-Learning,Robustness evaluation Digital Poster 3986 PRV Dose Sculpting for Hippocampal Sparing Whole Brain Radiotherapy in VMAT: A Dosimetric Comparison with TomoTherapy Deep S Pruthi, Arun Chairmadurai, Arpita Gupta, Shikha Halder, Shelley Hukku Radiation Oncology, Sir Ganga Ram Hospital, New Delhi, India Purpose/Objective: Hippocampal sparing whole brain radiotherapy (HS- WBRT) reduces neurocognitive decline by limiting hippocampal dose but its clinical implementation is often limited to helical TomoTherapy or non-coplanar VMAT platforms. We developed a novel coplanar three-arc method known as PRV-Sculpting Hippocampal Sparing (PRV-SHS). This study compared PRV-SHS with TomoTherapy plans of 1 cm and 2.5 cm jaw widths in the same cohort. Material/Methods: Ten patients undergoing HS-WBRT were replanned using three techniques: PRV-SHS on TrueBeam with HD MLC (using two coplanar whole-brain arcs + one hippocampal PRV sculpting arc), TomoTherapy 1 cm and TomoTherapy 2.5 cm jaw widths. Identical contours and optimization goals were applied. Dosimetric parameters included: PTV conformity index (CI), heterogeneity index (HI), D5%, D95%, D98%; hippocampal Dmean, D100%, Dmax; and OAR doses to lenses and optic nerves. Results: The results are summarized in the table. PRV-SHS achieved PTV coverage comparable to Tomo 1 cm and had significantly higher D95%, D98% compared to Tomo 2.5 cm (p<0.01) with superior conformity index and comparable dose homogeneity. PRV-SHS achieved hippocampal Dmean comparable with Tomo 1 cm (p=0.32 left, 0.29 right) and lower than Tomo 2.5 cm (p=0.01 left, 0.06 right); Dmax was comparable with Tomo 1 cm (p=0.41) and lower to Tomo 2.5 cm (p=0.006). However, D100% was higher than Tomo 1 cm but within acceptable constraints. Doses to other OARs like lenses and optic nerves were comparable across all techniques and within tolerance limits. Treatment time was significantly lower in the PRV-SHS
Figure 1: List of clinical goals for planning comparison. EX = external region. CI = conformity index. Results: PTVs were significantly larger for the EXT_L cohort (median: 1732/1643 cm ³ ) compared with INT_S (771/802 cm ³ ) and INT_M (1157/1256 cm ³ ) for left/right cases, respectively. Both the original and adapted models achieved comparable performance when applied to the OOD cohort: PTV coverage remained consistent across all groups, while median doses to organs at risk (OARs) increased proportionally with breast volume, as expected. A significant rise in hotspots within the breast PTV was observed for EXT_L patients compared with controls (V105% = 8.3/8.4% vs 7.0/6.3% and 7.8/6.5% for INT_S and INT_M, left/right respectively), though all remained within clinical limits. After minor model retuning to improve dose homogeneity, the updated model produced plans rated as equivalent or superior compared to the original model version in the qualitative assessment (left: 1.4 vs 1.4, 1.5 vs 1.9, 1.8 vs 2.5; right: 1.9 vs 2.3, 1.5 vs 1.9, 1.9 vs 2.5 for INT_S, INT_M, and EXT_L, respectively). The adapted right-sided model exhibited performance consistent with the original left-sided version, confirming the robustness and validity of the adaptation approach. Conclusion: The DL-based auto-planning models demonstrated strong generalization capability when applied to out- of-distribution patients, with only minor degradations in dose homogeneity—successfully mitigated through modest parameter retuning. These findings confirm the robustness of the model adaptation strategy for right-sided breast treatments and highlight that targeted adjustments can effectively enhance plan quality for patients outside the original training distribution.
Made with FlippingBook - Share PDF online