Comparación de modelos de inteligencia artificial para la detección de cardiopatías usando imágenes de resonancia magnética

In this study, three artificial intelligence techniques were compared for the binary detection (healthy/diseased) of cardiopathies using axial slices from cardiac magnetic resonance imaging. Using a dataset of 150 patients in NIfTI format, the images were preprocessed (normalization, rescaling to 128 × 128, RGB conversion, and data augmentation) and split on a per-patient basis using an 80/20 ratio. A Random Forest model with GLCM radiomic descriptors and first-order statistical features, a convolutional neural network (CNN), and a YOLOv8-based model adapted for binary classification were evaluated. The models were compared using accuracy, precision, recall, F1-score, and AUC, and explainability techniques (SHAP, Grad-CAM, Integrated Gradients, and occlusion sensitivity) were applied to validate the anatomical coherence of the predictions. Overall, the results indicate that pretrained deep learning approaches, such as YOLOv8, offer substantial advantages in terms of accuracy and interpretability, positioning them as a promising alternative for the development of intelligent decision-support systems for the diagnosis of structural cardiopathies.