Unsupervised Learning for Cardiac PET/MRI: A New Route to Mapping Myocardial Abnormalities

Lead

Cardiac imaging is becoming increasingly quantitative and multimodal, but turning that richness into clinically useful insight remains a major challenge. A new arXiv paper from researchers affiliated with Nantes Université, CHU Nantes and collaborators presents an unsupervised machine learning strategy for handling simultaneous cardiac PET/MRI data in patients with arrhythmogenic left ventricular cardiomyopathy.

This disease is a genetic myocardial disorder that is difficult to diagnose, partly because there is no clear gold-standard diagnostic criterion. PET/MRI can combine structural, tissue and metabolic information, yet this also creates a dense and heterogeneous data problem. Instead of training a supervised model to make a final diagnosis, the study focuses on organizing multimodal images into interpretable regional patterns that may reflect disease-related abnormalities.

Key points

• Clinical target: The study focuses on genetically diagnosed arrhythmogenic left ventricular cardiomyopathy, a condition where better characterization of phenotype and progression could be clinically valuable.
• Input data: The researchers used T1 and T2 maps, late gadolinium enhancement images and 18F-FDG-PET images from 99 patients.
• Two-step clustering: Each patient’s images were independently z-scored and summed into a single volume. These volumes were first clustered into supervoxels, then linked across patients using spectral clustering.
• Inter-patient groups: The process produced 32 groups of supervoxels shared across patients, creating a regional framework for comparison.
• Abnormality scoring: Each cluster and imaging modality received an abnormality score, allowing the system to visualize regions likely associated with disease.
• Automated reporting: The method generated both textual summaries and bullseye-style cardiac health reports.
• Evaluation: Compared with cardiac imagers’ assessments, the reports achieved a balanced accuracy of 0.76±0.04 in repeated nested cross-validation on patient data. On 167 numerical phantoms, balanced accuracy was at least 0.8.

Why it matters

The central contribution is not a single diagnostic classifier, but a framework for making complex PET/MRI data more systematic and comparable. In myocardial diseases, abnormal tissue patterns may vary by region and modality. Fibrosis, inflammation and metabolic changes can appear differently across images, making visual interpretation demanding even for experts. An unsupervised approach can help reveal recurring abnormal regions without requiring exhaustive manual labels.

The generated reports are best understood as decision-support tools rather than replacements for cardiac imaging specialists. A bullseye map can make regional findings easier to review, while automated text can summarize patterns in a consistent format. According to the paper, the abnormal clusters also closely matched visual observations, supporting their usefulness for highlighting varying degrees of fibrosis or inflammation.

Still, the study should be read as preliminary. It is based on a specific disease cohort and requires broader validation before clinical deployment. Larger, multi-center datasets and prospective testing would be needed to assess robustness, reproducibility and workflow integration. Even so, the work points to an important direction for medical AI: using machine learning not merely to label disease, but to transform multimodal imaging into structured, interpretable patient profiles.

Source: arXiv