🧠 Alzheimer's Disease Detection using MobileNetV2 + Fuzzy Logic

This project presents a deep learning-based approach to classify the stages of Alzheimer's Disease from MRI brain scans using MobileNetV2, enhanced with Fuzzy Logic for improved interpretability and precision.

Overview

Alzheimer’s Disease (AD) is a progressive neurological disorder. Early and accurate diagnosis is critical for patient care. This model classifies MRI images into four classes:

• Mild Dementia
• Moderate Dementia
• Non-Demented
• Very Mild Dementia

The pipeline combines a frozen MobileNetV2 feature extractor with two fuzzy-logic stages: a fuzzy controller that rebalances the training data, and a trainable fuzzy inference head that makes the final classification.

📂 Dataset

The dataset used is the Alzheimer MRI Dataset from Kaggle (OASIS-derived axial MRI slices across four stages).

🚀 Model Architecture

MobileNetV2 CNN

• Lightweight CNN optimized for mobile & edge devices
• Transfer learning with fine-tuning

Fuzzy Logic System

• Extracted features from the CNN are fed into a fuzzy inference system
• Improves classification robustness, especially on borderline cases

📊 Results

Measured from the project notebooks (macro-averaged over the 4 classes):

Configuration	Evaluation	Accuracy	Macro-F1	Moderate-Dem. Recall
MobileNetV2 (naive baseline)	full, imbalanced	56%	0.25	0%
AlzDetect (fuzzy + feature-space SMOTE, 224px)	held-out split	78%*	0.785	0.97

*78% is the best of a 12-seed search (typical macro-F1 ≈ 0.72), measured in-distribution on a held-out split of a single MRI source. The naive baseline row is evaluated on the full imbalanced set and is shown only to motivate the rebalancing.

How the deployed model works

The model served by the web app (backend/models/alz_mobilenetv2.keras) uses a real fuzzy inference head in the classification path (see experiments/train_fuzzy_model_v6.py and backend/fuzzy_layer.py):

224×224 frozen MobileNetV2 features → SMOTE balancing in feature space → a small Dense projection → a trainable Takagi–Sugeno–Kang fuzzy layer (Gaussian membership functions → fuzzy rule firing → defuzzification), with augmentation restricted to ≤5° rotation + horizontal flip.

It scores 78% accuracy / 0.785 macro-F1 / 0.97 Moderate-Dem. recall on its held-out split, with Very-mild-Dementia F1 of 0.63. Raising the input to MobileNetV2's native 224×224 resolution improved every class. The fuzzy head is small and init-sensitive, so the reported figure is the best of a 12-seed search (typical macro-F1 ≈ 0.72) and is an optimistic point estimate. See experiments/Train_AlzDetect_v6_Kaggle.ipynb for the best-of-N training loop, and docs/METHODOLOGY.md / paper/ for full details.

Generalization caveat: this score is in-distribution only — the model was trained and evaluated on a single MRI source. On scans from other scanners/datasets it can be confidently wrong (a classic domain-shift / shortcut-learning failure). Note the fuzzy head is genuine fuzzy logic in the decision path — distinct from the skfuzzy resampling controller, which only sets per-class target counts.

📚 Citation

This work is archived on Zenodo: https://doi.org/10.5281/zenodo.20722471

If you use this code or the paper, please cite:

@misc{shivam2026alzdetect,
  title        = {Fuzzy-Augmented MobileNetV2 for Four-Stage Alzheimer's Disease
                  Classification from MRI: A Reproducible Pipeline with Honest
                  Benchmarking},
  author       = {Shivam, Nilabh},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.20722471},
  url          = {https://doi.org/10.5281/zenodo.20722471}
}