Recurrence-specific supervised graph clustering for subtyping Hodgkin Lymphoma radiomic phenotypes.

The prediction at baseline of patients at high risk for therapy failure or recurrence would significantly impact on Hodgkin Lymphoma patients treatment, informing clinical practice. Current literature is extensively searching insights in radiomics, a promising framework for high-throughput imaging feature extraction, to derive biomarkers and quantitative prognostic factors from images. However, existing studies are limited by intrinsic radiomic limitations, high dimensionality among others. We propose an exhaustive patient representation and a recurrence-specific multi-view supervised clustering algorithm for estimating patient-to-patient similarity graph and learning recurrence probability. We stratified patients in two risk classes and characterize each group in terms of clinical variables.