3D-MobiBrainNet Alzheimer's AI

Overview

This peer-reviewed research implements volumetric brain MRI analysis to distinguish Alzheimer's disease, mild cognitive impairment and cognitively normal states. Rather than analyzing isolated 2D slices, the network processes three anatomical planes — axial, coronal and sagittal — through separate feature branches, then fuses them to preserve spatial context while staying computationally efficient.

Purpose — Why This Research Matters

Most Alzheimer's imaging models only do binary classification (disease vs. no disease) from single 2D slices, throwing away 3D spatial context that matters for early staging. Catching mild cognitive impairment — not just full-blown disease — is what actually enables timely intervention. 3D-MobiBrainNet was built to classify all three clinically relevant stages from true volumetric MRI, while staying lightweight enough for practical, real-world inference.

Animated Architecture Flow

1

3D MRI Volume

Full volumetric brain scan loaded via NiBabel.

2

Multi-Plane Extraction

Axial, coronal and sagittal views extracted.

3

Feature Branches

Depth-wise separable convs per plane.

4

Feature Fusion

Recalibration merges the three branches.

5

Classification Head

Multi-class prediction across disease states.

6

Diagnosis Output

AD / MCI / Cognitively Normal, with confidence.

Key Features

True 3D volumetric processing of brain MRI rather than isolated 2D slices.

Multi-plane feature fusion across axial, coronal and sagittal views.

Depth-wise separable convolutions for reduced computational complexity.

Feature recalibration emphasizing the most informative representations.

Multi-class classification distinguishing three disease states simultaneously.

Efficient architecture — 34.1M parameters, far fewer than comparable models, with 5-fold cross-validation.

Benefits

Earlier Staging

Distinguishes MCI from full Alzheimer's, not just disease vs. healthy.

True 3D Context

Multi-plane fusion keeps spatial context 2D-slice models lose.

Lightweight & Fast

Far fewer parameters than comparable 3D CNNs, easing real deployment.

Peer-Reviewed

Published in Ain Shams Engineering Journal (Elsevier) — independently validated.

Results

0%

Accuracy

0%

Precision

0%

Recall

0%

F1-Score

0%

AUC

0M

Parameters

Evaluated on the Alzheimer's Disease Neuroimaging Initiative (ADNI) 3D MRI dataset.

Publication

Journal: Ain Shams Engineering Journal (Elsevier) · Year: 2025 · Volume: 16 · Issue: 11 · Article: 103714
DOI: 10.1016/j.asej.2025.103714 · Open Access, Peer-Reviewed

Technology Stack

Python 3.10+ PyTorch 3D CNN NiBabel Scikit-learn Matplotlib Jupyter

Explore the Repository Read the Paper