Revisiting Multi-Nodal Radiomics with Advanced Feature Learning for Lymphoma Classification: A Multi-Center Study

We develop and evaluate a patient-level framework for discriminating Hodgkin lymphoma (HL) from non-Hodgkin lymphoma (NHL) on baseline 18F-FDG PET by aggregating radiomic descriptors from a variable number of involved lymph nodes. Each patient is represented as an unordered set of lesion-level radiomic feature vectors, optionally augmented with demographic and spatial information. We compare statistical pooling with advanced set-based aggregation models, including Transformers, Deep Sets, multiple instance learning, Set Transformers, and graph-based learning. The models are trained on one clinical center and externally evaluated on an independent center, including unsupervised test-time adaptation to address domain shift. Combining radiomic, demographic, and spatial features improves discrimination, while test-time adaptation further improves cross-center generalization.

Physics in Medicine & Biology, 71(12), 125014, 2026