A novel unsupervised machine learning clustering strategy to identify PET/MR biomarkers in arrhythmogenic cardiomyopathy

Motivation: Diagnosing arrhythmogenic cardiomyopathy (AC) is challenging without gold-standard criteria. LGE, T1/T2 mappings, and PET imaging offer complementary insights on fibrosis and inflammation. Goal(s): This study aims to demonstrate the potential of simultaneous PET/MR and inter-patient data linkage to discover novel regional markers of AC. Approach: Two-step clustering was applied to multimodal images of AC patients. Supervoxels were extracted from each patient, before being clustered to thirty-two inter-patient groups. Patient's health reports were generated and compared to cardiac imagers' reports using balanced accuracy (BA). Results: Clustering reports accurately represented the proportions of hyper-signal combinations per patient, while identifying most cardiac imagers observations (BA=0.76). Impact: A two-step multimodal PET/MR unsupervised clustering method combining supervoxel extraction and inter-patient clustering was developed, enabling robust identification, visualization, and quantification of abnormal regions in arrhythmogenic cardiomyopathy patients. It provides an encouraging step toward identifying prognostic clusters and patient profiles.