An explainable deep learning framework for differentiating idiopathic normal pressure hydrocephalus from Alzheimer’s disease using brain MRI

Background and objectives Idiopathic normal pressure hydrocephalus (NPH) and Alzheimer’s disease (AD) may present with overlapping clinical and neuroimaging features, complicating differential diagnosis. This study evaluated an explainable, attention-enhanced deep learning framework for distinguishing NPH from AD using two-dimensional brain magnetic resonance imaging (MRI). Methods The dataset comprised 96 patients (48 NPH and 48 AD) and 295 MRI images representing axial, sagittal, and coronal anatomical planes. The reported patient-level partition comprised 68 patients (208 images) for training, 14 patients (42 images) for validation, and 14 patients (45 images) for testing. Five ImageNet-pretrained convolutional neural network architectures–EfficientNetB2, ResNet152, VGG19, DenseNet201, and Xception–were evaluated with the Convolutional Block Attention Module (CBAM). EfficientNetB2 without CBAM was included as a baseline. Classification performance was assessed using standard performance metrics, confidence intervals, and paired statistical comparisons. Grad-CAM was used to visualize model activation patterns. Results EfficientNetB2 with CBAM achieved the highest observed image-level classification accuracy of 93.33% (42/45; patient-cluster bootstrap 95% CI: 86.05–100.00%) and a receiver operating characteristic area under the curve (ROC-AUC) of 0.9722 (95% CI: 0.9127–1.0000). For NPH classification, sensitivity and specificity were 90.48% and 95.83%, respectively. The corresponding EfficientNetB2 baseline achieved 88.89% accuracy, an exploratory difference of 4.44 percentage points; the paired difference was not statistically significant (exact McNemar test, p = 0.500). Grad-CAM visualizations revealed varying activation patterns across the six configurations without establishing clinical validity. Conclusion Attention-enhanced convolutional neural networks demonstrated promising image-level performance for differentiating NPH from AD on the evaluated MRI dataset. Nevertheless, the limited patient sample, absence of patient-level prediction aggregation, and lack of independent external validation warrant cautious interpretation. Further validation in larger and heterogeneous clinical cohorts is required before clinical application.