Cardiomyopathy classification from MRI using attention-based deep learning models

Abstract

Cardiomyopathies are a leading cause of heart failure, and early detection is critical for effective patient management. The growing availability of cardiac magnetic resonance imaging (MRI) generates large volumes of data, creating challenges for accurate and efficient analysis. In this study, we propose a novel methodology that integrates radiomic features, capturing statistical and texture-based information from MRI scans, with deep features extracted from a classical convolutional neural network (ResNet50). To enhance both interpretability and feature integration, attention mechanisms—including self-attention and Squeeze-and-Excitation blocks—are employed to selectively weight the contributions of each modality. Our approach achieves up to 78% classification accuracy and demonstrates significant gains in attention-based explainability metrics, particularly emphasizing the importance of radiomic features. These findings indicate that combining radiomic and deep features with attention mechanisms provides a robust framework for cardiomyopathy classification, with radiomic information playing a dominant role in guiding model decisions.