ESTRO 2025 - Abstract Book

S3376

Physics - Machine learning models and clinical applications

ESTRO 2025

Results: MMDP achieved accurate dose predictions, with DVH score of 1.94 Gy, Dose score of 1.14 Gy, and prediction errors for functionally weighted mean lung dose (fMLD) of 0.10±0.57 Gy. In contrast, MMDP trained through naive multi modality joint training and a commonly-used anatomy-guided dose prediction model significantly overestimated fMLD by 0.40±0.58 Gy ( p <0.001) and 0.30±0.60 Gy ( p <0.001) respectively. DL-FLART plans demonstrated comparable fMLD compared to high-quality MO-FLART plans (6.57±3.05 Gy versus 6.60±3.18 Gy, p =0.334) and lower fMLD compared to low-quality MO-FLART plans (7.37±3.67 Gy versus 7.81±4.05 Gy, p <0.001). DL FLART plans exhibited lower fMLD (5.91±3.04 Gy versus 6.15±3.03 Gy, p <0.001) on SPECT V dataset and comparable fMLD (7.27±3.20 Gy versus 7.12±2.71 Gy, p =0.099) on SPECT Q dataset respectively compared to MA-FLART plans. Compared to MA-ConvRT plans, DL-FLART plans effectively achieved lung dose painting as demonstrated by Figure 2, and reduced fMLD by 0.80 Gy ( p <0.001) and 0.46 Gy ( p <0.001) on SPECT V and Q datasets respectively.