ESTRO 2025 - Abstract Book

S3399

Physics - Machine learning models and clinical applications

ESTRO 2025

Logistic regression based NTCP-models for dysphagia served as a reference and were refit at each time-point using dosimetric and clinical features. Inputs for deep learning models were pre-treatment 3D dose, CT-images, segmentation of swallowing structures, and clinical features. All deep learning models were based on a Resnet architecture, providing automatically learned dose and CT-related features, which were combined with clinical features in the fully-connected layers to generate the final predictions. For multi-label and sequential approaches, regular linear layers were modified into multi-headed (simultaneous) and sequentially ordered (one-by-one) predictions (Figure 1). Models were trained using mean binary-cross-entropy across all time-points as loss function with Adam as optimizer.

Results: All approaches reached good results for each time-point on the independent test set, with a median AUC of 0.71 [inter-quartile range(IQR): 0.1] for logistic regression and median AUC of 0.80 [IQR: 0.12] with deep learning (Table 1), demonstrating that deep learning-based approaches generally outperform conventional NTCP models. Multi label and sequential deep learning approaches showed similar performance (AUC=0.76 [IQR: 0.14], AUC = 0.78 [IQR: 0.15] respectively). However, both were generally outperformed by the simpler per time-point deep learning method (AUC = 0.83, [IQR: 0.08]).