ESTRO 2025 - Abstract Book

S3446

Physics - Machine learning models and clinical applications

ESTRO 2025

The mean absolute errors between PRs calculated using measured and GAN-predicted dose maps were 3.6% (SD = 2.5%) and 5.6% (SD = 3.5%) for the 3%G, 2 mm and 2%G, 2 mm criteria, respectively. Average mGI differences were 0.05 (SD = 0.06, median = 0.08) and 0.06 (SD = 0.07, median = 0.06), respectively. As shown in Figure 2, PRs from GAN-predicted maps were generally underestimated, while mGIs were overestimated. Discrepancies were largest in plans with PR(3%G, 2 mm) < 90%, a range typically linked to PSQA failures [2]. Notably, these cases were underrepresented in the training dataset.