Revisiting multi-nodal radiomics with advanced feature learning for lymphoma classification: a multi-center study.

Area of Science:

• Medical Imaging
• Radiomics
• Machine Learning

Background:

• Distinguishing Hodgkin lymphoma (HL) from non-Hodgkin lymphoma (NHL) is crucial for treatment.
• ¹⁸F-FDG PET imaging offers valuable diagnostic information.
• Radiomics extracts quantitative features from medical images, showing potential in cancer subtyping.

Purpose of the Study:

• To develop and evaluate a patient-level framework for discriminating HL from NHL using baseline ¹⁸F-FDG PET.
• To aggregate radiomic descriptors from multiple lymph nodes in a multi-center setting.
• To assess the effectiveness of advanced machine learning models and test-time adaptation for improved cross-center generalization.

Main Methods:

• Modeled patients as sets of lesion-level PET radiomic feature vectors.
• Extracted 107 radiomic features (first-order, texture, shape) using PyRadiomics.
• Compared statistical models with set-based aggregation models (Transformer, Deep Sets, etc.), incorporating demographic and spatial data.
• Trained and validated models on multi-center data, evaluating direct transfer and unsupervised test-time adaptation (TTA).

Main Results:

• Combining radiomics with demographic and spatial features improved HL vs. NHL discrimination.
• Best deep aggregation models achieved high performance on the training center (ROC-AUC up to 0.91).
• Unsupervised TTA (e.g., TENT) significantly improved cross-center generalization on the external validation cohort, enhancing Transformer performance (ROC-AUC from 0.81 to 0.87).

Conclusions:

• Multi-nodal PET radiomics, augmented with demographic and spatial information, shows promise for patient-level HL vs. NHL prediction.
• The study highlights the need for robustness in multi-center PET radiomics.
• Test-time adaptation is a valuable strategy for mitigating cross-center variability and improving model generalizability.