Cardiomyopathy detection using ensemble-based segmentation and radiomics
T Tsampras, A Antonopoulos, T Karamanidou, G Kalykakis, K Tsioufis, C VlachopoulosAbstract
Background
Diagnosis of cardiomyopathies often relies on overt phenotypic manifestations, leading to delayed clinical recognition and management. There is a growing need for scalable, population-level screening tools capable of detecting subclinical myocardial disease.
Purpose
The purpose of this study was to develop a Machine Learning (ML) model to automatically segment the left ventricular myocardium from CT scans and estimate the probability of underlying myocardial disease using radiomic feature analysis.
Methods
An ensemble ML model was developed to automatically segment the left ventricular (LV) myocardium from CT images. The model was trained on 60 CT scans (~12,000 images) acquired using multiple CT protocols, and externally validated on 10 independent scans. Four high-performing ML models were combined into an ensemble model. Radiomic features were extracted from both manual and automatic LV segmentations on a separate cohort of 100 CT scans: 50 healthy individuals and 50 patients with myocardial disease (i.e., patients with known cardiomyopathy diagnosed and monitored in specialized cardiomyopathy Units). Feature stability, cross-variation, and correlation analyses were performed, followed by disease classification using a random forest model. Key predictive features were identified, and model performance was evaluated.
Results
The four best-performing models (Unet++, ED w/ASC, FPN, and TresUNET) were combined to form an Ensemble model, achieving a final DICE score of 0.882 after hyperparameter optimization. External validation yielded a DICE score of 0.907. Radiomic analysis identified 15 key features predictive of myocardial disease. The classification model demonstrated good diagnostic performance, achieving an AUC of 0.80 in automatically generated segmentations and 0.85 in manual segmentations, indicating robust and consistent disease prediction across segmentation methods.
Conclusions
This study presents a fully automated CT-based framework for LV myocardial segmentation and radiomic phenotyping that estimates the probability of underlying myocardial disease. The model demonstrates generalizability across different CT protocols and highlights the potential role of AI-driven CT analysis for early, non-invasive cardiomyopathy screening at a population level.Segmentation examples-Disease predictionFor image description, please refer to the figure legend and surrounding text.Automated disease classification modelFor image description, please refer to the figure legend and surrounding text.