Abstract :

Deep Learning Models For Alzheimers Disease (AD) Detection From Magnetic Resonance Imaging (MRI) Usually Report High Accuracy, But They Are Trained On Thousands Of Labelled Scans, Run On GPUs, And Are Often Evaluated With Slice-level Splits That Leak Information Between Training And Testing. This Paper Presents ElasticAD, A Lightweight And Resource-adaptive Framework Designed For The Realistic Case Of A Few Hundred Labelled Subjects And Low-cost Hardware. ElasticAD Converts Each T1-weighted Volume Into 2.5D Inputs Made Of Three Adjacent Axial Slices, Learns General Brain-tissue Features From Unlabelled Scans Through Selfsupervised Contrastive Pre-training, And Then Trains A Compact Student Network Built From Ghost Modules And Coordinate Attention Under The Guidance Of A Larger ConvNeXt-Tiny Teacher. A Class-balanced Loss Handles The Shortage Of Demented Subjects. To Adapt To Different Devices, The Trained Student Is Compressed By Budget-aware Channel Pruning And Fitted With A Confidence-based Early-exit Branch, And Slice Predictions Are Fused Into One Decision Per Subject. On 198 Labelled OASIS-1 Subjects Classified As Cognitively Normal, Very Mild Dementia Or Mild-to-moderate Dementia, And Evaluated With Strict Subject-level Five-fold Crossvalidation, ElasticAD Reached 85.3% Balanced Accuracy And A Macro F1-score Of 0.85 With Only 0.84 Million Parameters And 70 MFLOPs Per Input. It Outperformed ResNet18, DenseNet121, MobileNetV3, ShuffleNetV2 And EfficientNet-B0, And It Matched Its Own 28-million-parameter Teacher. With Only 25% Of The Labelled Subjects It Still Achieved 77.8% Balanced Accuracy, 11.3 Points Above EfficientNet-B0. On A Raspberry Pi 4 The Full Model Needs 12.1 Ms Per Input And The Pruned 30% Variant Only 4.2 Ms, Showing That Reliable AD Screening Is Possible Without Large Datasets Or GPUs.

Published:

16-10-2026

Issue:

Vol. 26 No. 10 (2026)

Page Nos:

61 - 68

Section:

Articles

License:

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

How to Cite