<![CDATA[Stroke is a major cause of neurological disability, and rapid, accurate diagnostic support remains essential. Because MRI interpretation still depends heavily on radiologist expertise, automated classification models are needed to support Computer-Aided Diagnosis (CAD). This study compares ConvNeXt V2 and Vision Transformer (ViT) for classifying stroke lesions in MRI images, using 7,463 images (4,423 Stroke; 3,040 Normal) integrated from two public Kaggle datasets. The pipeline covered dataset integration, preprocessing, augmentation, transfer learning with full fine-tuning, and evaluation on a held-out 20% test set using accuracy, precision, recall, specificity, F1-score, and AUC-ROC, supported by Grad-CAM and attention-rollout saliency analysis. ConvNeXt V2 achieved 94.98% accuracy, 95.81% precision, 95.71% recall, 93.91% specificity, 95.76% F1-score, and 99.25% AUC, outperforming ViT (92.77% accuracy, 94.30% precision, 93.45% recall, 91.78% specificity, 93.87% F1-score, 98.54% AUC) across all metrics. The largest gap appeared in accuracy and recall, and a paired McNemar test confirmed the difference is statistically significant (χ² = 14.42, p < 0.001). These findings support ConvNeXt V2 as a promising architecture for MRI-based stroke CAD systems, while highlighting the continued need for external clinical validation. ]]>
Copyrights © 2026
