Grey wolf optimization-driven convolutional neural network for robust cardiac MRI-based heart disease classification under class imbalance

Heart disease diagnosis through Cardiac Magnetic Resonance Imaging (MRI) is a vital clinical diagnostic technique that enables the non-invasive evaluation of cardiac structure and identification of cardiovascular diseases. Nevertheless, the automation of such a task using deep learning faces some critical problems in terms of limited data size, class imbalance, and manual tuning of hyperparameters, which might hinder the accuracy and reliability of the process. Therefore, this paper introduces a novel algorithm called GWO-CNN for heart disease classification based on cardiac MRI. In this algorithm, structured pre-processing, contrast enhancement, augmenting-based class balancing, CNN-based spatial features extraction, and GWO-based hyperparameters selection will be used for the classification process. Cardiac MRIs are preprocessed through normalization, Gaussian filtering, contrast-limited adaptive histogram equalization, cropping, and resizing improve anatomical clarity and generate standardized image inputs. Moreover, the classifier model consists of a CNN model to identify between normal and abnormal heart MRIs and Grey Wolf Optimization algorithm to find optimal values of hyperparameters (learning rate, dropout rate, batch size, configuration of convolutional filters, and dense layer size). Experimental results, obtained via patient-level five-fold cross-validation repeated across five random seeds (25 runs in total), demonstrated that the GWO-CNN framework attained a mean accuracy of 98.30% ± 0.42% (95% CI: 98.13–98.47%), a mean F1-score of 97.50% ± 0.46%, and a mean AUC of 98.60% ± 0.39%, with a representative single-fold configuration reaching 98.42% accuracy, 97.86% precision, 97.31% recall, a 97.58% F1-score, a specificity of 98.91%, a Dice score of 97.52%, a Jaccard similarity index of 95.23%, a Matthews correlation coefficient (MCC) of 96.41%, and an area under the curve (AUC) of 98.74%. Paired statistical testing (t-test and Wilcoxon signed-rank, p < 0.001) confirmed that the improvements attributable to preprocessing, GAN-based augmentation, and GWO-based tuning, as well as the improvement over baseline CNN, ResNet-based, and transfer-learning-based comparators, are each statistically significant. The results demonstrate the potential of the proposed GAN-assisted GWO-CNN framework for cardiac MRI classification under limited and class-imbalanced data conditions. A preliminary external evaluation on the independent ACDC dataset, performed without any retraining, yielded 92.00% accuracy, a moderate decrease from the internal cross-validated mean consistent with real but non-catastrophic distributional shift. However, because the primary evaluation is based on a relatively small single public dataset and external validation was limited to one additional cohort assessed once, these findings should be interpreted as preliminary proof-of-concept evidence rather than confirmation of clinical applicability. Further validation using larger and independent multi-center cardiac MRI cohorts is required before the proposed framework can be considered for clinical deployment.