AttentionMS-Net: Alzheimer's Disease Classification from MRI
Osman Yildiz
AttentionMS-Net
An Attention-Enhanced Multi-Scale Framework for Alzheimer's Disease Classification with Subject-Level Validation
Osman Yildiz¹ and Abdulhamit Subasi¹
¹ Department of Information Science and Technology, University at Albany, State University of New York, Albany, NY 12222, USA
Overview
This repository contains the complete experimental pipeline for our paper on Alzheimer's disease classification from 2D MRI slices. The study makes three key contributions:
Data leakage quantification: We demonstrate that image-level splitting inflates accuracy by ~19 percentage points compared to proper subject-level splitting (99.9% vs 80.8%).
Three-level evaluation hierarchy: slice-level → subject-level → subject-level with prediction aggregation, each progressively reducing bias.
CNN vs. Transformer comparison: Systematic evaluation of AttentionMS-Net (end-to-end CNN with CBAM + multi-scale fusion) against frozen Swin Transformer + ML classifiers under identical 10-fold subject-level cross-validation.
Key Results
Model Accuracy F1 Macro AUC-ROC Sensitivity Specificity AttentionMS-Net (slice) 0.805±0.047 0.726±0.057 0.866±0.056 0.609±0.146 0.861±0.069 AttentionMS-Net (subject) 0.824±0.070 0.753±0.106 0.889±0.055 0.653±0.243 0.876±0.066 Swin+Stacking (slice) 0.815±0.045 0.741±0.050 0.879±0.043 0.618±0.064 0.872±0.051 Swin+Stacking (subject) 0.841±0.050 0.784±0.068 0.904±0.042 0.701±0.123 0.883±0.045
Wilcoxon signed-rank test: p = 1.000 — no statistically significant difference between the two approaches. AttentionMS-Net provides gradient-based interpretability (Grad-CAM++) that frozen pipelines cannot.
Dataset
We use the OASIS-1 cross-sectional MRI dataset (Marcus et al., 2007), preprocessed as 2D axial slices:
347 subjects (271 Non-Demented, 76 Demented)
86,437 slices total (~250 slices per subject)
Binary classification: Non-Demented vs. Demented
The dataset is available via Kaggle. Download and place in
data/raw/.Project Structure
Installation
Reproducing Results
1. Data Preparation
2. CNN Experiments (AttentionMS-Net)
3. Swin Transformer Experiments
4. Evaluation
5. Explainability
Running on SLURM (DGX Cluster)
Hardware
All experiments were conducted on:
GPU: NVIDIA A100-SXM4-80GB (DGX cluster)
CPU: AMD EPYC 7742 (for ML classifiers)
Software: Python 3.10, PyTorch 2.1, CUDA 12.1
Citation
If you use this code, please cite:
License
This project is licensed under the MIT License — see LICENSE for details.
Authors
Osman Yildiz — Department of Information Science and Technology, University at Albany, SUNY (oyildiz@albany.edu)
Abdulhamit Subasi — Department of Information Science and Technology, University at Albany, SUNY (asubasi@albany.edu)
OASIS dataset: Marcus et al., 2007
University at Albany Research IT for DGX cluster access
Like this project
Posted Sep 5, 2026
Alzheimer's MRI classifier with subject-level 10-fold CV (AUC 0.89). Quantified ~19 pp accuracy inflation from image-level leakage. PyTorch, Grad-CAM++.
Likes
0
Views
1
