ESTRO 2025 - Abstract Book

S3412

Physics - Machine learning models and clinical applications

ESTRO 2025

The dataset was divided into training (n=816), validation (n=272), and test (n=121) subsets to develop a VGGNet16 based model architecture.

Results: The results highlight the model’s discriminative power and good generalization on unseen data with AUC=0.89, SENS=0.8, and SPEC=0.87 on the test set. Results in terms of AUC, PREC, SPEC, SENS, confusion matrices, and ROC curves for all the datasets are shown in Figure 2 . Our strategy focuses on minimizing false positives – plans wrongly classified as “pass” that would fail – even if it leads to an increased number of false negatives.