Efficient neuroanalytic local–global learning with dual-branch CNN–lightweight Swin Transformer for Alzheimer’s disease analysis

Alzheimer’s disease (AD) is one of the main causes of dementia, where early detection is critical for postponing the progression of the disease and timely treatment. Recent computer-aided diagnosis using deep learning with structural MRI has shown optimal performance; however, the rising number of model parameters and high computational requirements make the deployment challenging in resource-limited environments. This study proposes an efficient deep learning framework named “NeuroCST” based on a dual-branch network architecture incorporating an enhanced EfficientNet-based CNN and a Lightweight Swin Transformer working independently to capture local and global information of AD. The enhanced Efficient Convolution branch extracts fine-grained morphological features using depthwise separable convolution and residual learning for texture capturing. Consequently, the customized SwinLite Transformer branch models global contextual dependencies using Efficient SwinLite Transformer blocks with attention, shifted-window partitioning, and MLP. These dual-branch diverse feature vectors are then concatenated and processed by a spatial attention mechanism, NeuroCST-SA, for optimal pixel selection to reduce redundancy and discriminate minor contrast and texture inter-class variation. The proposed NeuroCST is evaluated on the public Kaggle AD dataset, which contains 6,400 original and 33,984 augmented MRI images (40,384 in total) in four classes: Mild Demented, Moderate Demented, Non-Demented, and Very Mild Demented. The NeuroCST-SA performs better on diverse experimental setups, both augmented and original, than existing CNNs and ViTs, achieving 98.42% accuracy, 98.50% sensitivity, 98.60% precision, and an F1-score of 98.55% (mean, with a standard deviation of 2.49% across holdout cross-validation). The base NeuroCST model has 15.3 million parameters and 5.2 GFLOPs, and the NeuroCST-SA variant has 14.8 million parameters, 4.9 GFLOPs, and a processing time of 11.4 ms per image. The performance demonstrates the effective integration of CNN and transformer in capturing diverse local–global features for robust AD classification.