Attention-Augmented EfficientNetV2B0 for Multi-Class Cardiac Disease Classification from Cine MRI

Abstract

This study presents a pipeline of deep learning for multiclass diagnosis of cardiac disease from cine MRI that combines three significant innovations: an attention-augmented EfficientNetV2B0 backbone for enhanced spatial discrimination of cardiac anatomy, domain-specific preprocessing using CLAHE and best slice selection to enhance prominent myocardial features, and a patient-level ensemble strategy that aggregates slice-wise predictions into robust diagnostic outputs. The model accounts for the volumetric and heterogeneous nature of cardiac MRIs, unlike traditional perslice approaches. We evaluated our system on the MICCAI 2017 ACDC dataset for five classes dilated cardiomyopathy (DCM), hypertrophic cardiomyopathy (HCM), myocardial infarction (MINF), abnormal right ventricle (ARV), and normal (NOR) with a )90.0 %( overall test accuracy and macro F1-score of 0.9013. Performance per class was extremely high in ARV and HCM, with precision and recall of over 0.90. Cross-validation also confirmed the stability of the model with a mean accuracy of 69.6 % ± 1.1 and an F1-score of 67.9% ± 2.0. The model exploited transfer learning with partial fine-tuning as well as attention’s saliency maps, achieving both generalizability and interpretability clinically. By smoothing patient level predictions and regulating model attention toward radiological expectations, the system provides a more consistent and trustworthy diagnosis and has the potential to reduce cardiac triage time by as much as 40 %. The essence of the challenge is still the detection of MINF on the basis of faint tissue biomarkers. Overall, our contribution enhances computational cardiology with a realizable, explainable, and highly accurate method for automatic diagnosis.