A novel unsupervised machine learning strategy to handle multimodal cardiac PET/MRI data

TL;DR — A preliminary study proposing an unsupervised clustering pipeline to jointly analyze multimodal cardiac PET/MRI data (T1/T2 maps, LGE, 18F-FDG-PET) and auto-generate health reports for arrhythmogenic left ventricular cardiomyopathy patients. It matters because it offers a systematic, cross-modal way to characterize myocardial heterogeneity for a disease that lacks gold-standard diagnostic criteria.

• Method: Per-patient images are z-scored and summed into a single volume, segmented into supervoxels, then grouped across 99 patients into 32 inter-patient clusters via spectral clustering (a two-step clustering approach).
• Outputs: Each cluster/modality gets an "abnormality" score used to flag disease-associated regions and produce automated textual and bullseye health reports.
• Validation: Reports matched cardiac physician assessments with balanced accuracy of 0.76 ± 0.04 (repeated nested cross-validation on patients) and BA ≥ 0.8 on 167 numerical phantoms; flagged clusters aligned with visual fibrosis/inflammation observations.
• Note: Framed as a methodological, proof-of-concept study on a small cohort rather than a validated clinical tool.