earticle

영어

Early diagnosis of dementia, such as Alzheimer's disease, is clinically crucial, and deep learning-based brain MRI analysis has gained significant attention. However, acquiring large-scale datasets remains challenging in medical imaging, necessitating learning from small datasets. This study quantitatively compares the dementia classification performance of CNN-based ResNet-18 and various Transformer architectures (Swin Transformer, Vision Transformer, UNETR) using the OASIS-3 brain MRI dataset (3,428 samples) and systematically evaluates their suitability for small-scale medical data. Methods: Data comprising 2,943 normal (85.8%) and 485 dementia (14.2%) cases were split 80:20, with 5-fold cross-validation. Focal Loss was applied to mitigate class imbalance, and metrics including Sensitivity, Specificity, Balanced Accuracy, and AUC-ROC were evaluated. Results: ResNet-18 achieved the most balanced performance with Sensitivity of 79.38%, Balanced Accuracy of 77.38%, and AUC-ROC of 85.40%. Transformer models showed distinctly different patterns: Swin Transformer (Sensitivity 42.27%, AUC 81.78%) exhibited normal-class bias, Vision Transformer (Sensitivity 22.68%, AUC 47.89%) nearly failed to learn with pure global attention, and UNETR (Sensitivity 88.66%, AUC 73.00%) achieved highest sensitivity but severely low specificity (41.77%). ResNet-18 also demonstrated superior learning efficiency (parameters: 33M, training time: 2.1h) compared to Transformers (Swin: 60-70M/3.1h, ViT: 18M/3.0h, UNETR: 10- 12M/4.0h). This study confirms that CNN's inductive bias and structural efficiency are more effective than Transformer's global attention or hybrid approaches in small-scale brain MRI datasets, with ResNet-18 proving most suitable for dementia screening due to balanced sensitivity-specificity trade-off.