Label-Efficient and Edge-Ready Alzheimer’s Disease Staging from MRI using Knowledge Distillation and Adaptive Inference

Alzheimer's disease (AD) is a progressive neurodegenerative disorder, and early detection from structural MRI can slow decline through timely clinical intervention. Deep convolutional neural networks have shown strong accuracy for AD staging, but they need large annotated datasets and high-end GPUs, which most hospitals and research labs in developing regions do not have. This paper proposes LRA-Net, a Lightweight and Resource-Adaptive deep learning framework that classifies brain MRI slices into four stages: Non-Demented, Very Mild Demented, Mild Demented and Moderate Demented. LRA-Net combines depthwise separable convolutions, inverted residual blocks and a channel-spatial attention module, and it scales its width and input resolution to match the memory and compute budget of the target device. To counter the small and imbalanced dataset, the framework uses skull-stripping, intensity normalization, class-aware augmentation, transfer learning from ImageNet and a focal loss. Experiments on the 6,400-image Kaggle Alzheimer's MRI dataset show that LRA-Net reaches 97.8% accuracy and a 0.97 macro F1-score with only 1.9 million parameters and 0.31 GFLOPs. It matches or exceeds heavier models such as VGG16 and ResNet50 while using over 13 times fewer parameters than ResNet50, and it runs at 18 ms per image on a CPU. Grad-CAM maps confirm that predictions focus on the hippocampus, ventricles and cortical regions known to atrophy in AD. The results show that accurate AD screening is feasible on low-cost hardware with limited data.