A novel unsupervised machine learning clustering strategy to identify PET-MR biomarkers of myocardial inflammation in arrhythmogenic cardiomyopathy

Abstract Background Arrhythmogenic cardiomyopathy (AC) is a genetic myocardial disease characterized by fibro-fatty myocardial replacement detected by late gadolinium enhancement (LGE) on magnetic resonance (MR). Myocardial inflammation (MI) can be an associated condition of which non-invasive diagnosis by MR and/or positron emission tomography (PET) imaging is challenging. Purpose To develop an unsupervised machine learning clustering approach for identifying hybrid PET/MR-derived phenotypes associated to MI in AC patients. Methods Consecutive AC patients carrying a pathogenic genetic variant were explored on an hybrid PET/MR system. Two-step clustering was applied to T1 and T2 maps, LGE and 18F-FDG-PET images (Figure 1). Left ventricular (LV) myocardial segmentation was performed semi-automatically on the LGE images. Images were z-scored, summed into a single volume, before being clustered into supervoxels using SLIC independently for each patient. Each supervoxel was assigned the median values of its voxels in each of the four imaging modalities. Supervoxels with extreme relaxation times were excluded. Data from all patients within the LV were then pooled and Box-Cox transformed. Thirty-two inter-patient groups of supervoxels were obtained by spectral clustering. An "abnormality" score or value was assigned to each cluster for each modality, calculated as the distance from its median compared to the healthiest cluster according to medical practice. A naive Bayesian network (BN, Fig.1) was defined with the occurrence of myocarditis before imaging the patient as central node, surrounded by binary clinical and clustering nodes. The BN was trained on all patients. Probabilities of each variable were inferred given that myocarditis had occurred or that it had not. The 2 cases were subtracted for each variable to determine which correlated most with myocarditis. The cluster features were reported to their clustering scores to relate to the PET/MR modalities. Results Fifty three patients were included, among which 8 had experienced a myocarditis before PET-MR imaging. While chest pain remained the feature most correlated with myocarditis, clusters showing abnormally high LGE and PET values had high probabilities of being associated with a history of myocarditis (Figure 2). On the contrary, clusters with the lowest PET abnormality scores were the less asociated with a probability of myocarditis. T1 and T2 seemed to be features of lower interest to predict the probability of myocarditis history. Conclusion We report a preliminary experience with a novel unsupervised machine learning clustering strategy of FDG-PET/MR biomarkers in AC patients at risk of inflammatory complications. LGE combined with PET seemed to be the most promising features to identify patients with a history of myocarditis. Further studies are needed to confirm the potential of this method to detect MI in AC.Figure 1.Methods Figure 2.Results