Alzheimer's Disease Detection using CNN

A deep learning system that classifies brain MRI scans into four stages of Alzheimer's-related cognitive decline using a lightweight Convolutional Neural Network, paired with a web-based clinical decision-support interface and an AI chatbot for patient-friendly result interpretation.

Classification Stages

• 🟢 No Impairment — Cognitively healthy
• 🟡 Very Mild Impairment — Earliest signs of decline
• 🟠 Mild Impairment — Moderate atrophy patterns
• 🔴 Moderate Impairment — Advanced neurodegeneration

Results

Metric	Value
Test Accuracy	98.83%
Inference Speed	~2.4 ms/image
Epochs Trained	94 (EarlyStopping)
Model Size	~10 MB

Project Structure

alzheimer_detection/
├── app.py                  # Flask web app + chatbot API
├── train.py                # Model training pipeline
├── test.py                 # Model evaluation & metrics export
├── generate_figures.py     # Academic figure generation (11 figures)
├── setup_dataset.py        # Kaggle dataset downloader
├── trained_model.keras     # Saved CNN model
├── requirements.txt        # Python dependencies
├── LICENSE
│
├── static/                 # Frontend assets
│   ├── style.css
│   └── script.js
│
├── templates/
│   └── index.html          # Web UI with chatbot widget
│
├── data/                   # Generated JSON (gitignored)
│   ├── training_history.json
│   └── evaluation_results.json
│
├── figures/                # Generated paper figures (gitignored)
│   ├── training_curves.png
│   ├── confusion_matrix.png
│   ├── cnn_architecture.png
│   ├── system_pipeline.png
│   ├── metrics_table.png
│   ├── sample_scans.png
│   ├── dataset_distribution.png
│   ├── per_class_accuracy.png
│   ├── flowchart.png
│   ├── prototype_ui.png
│   ├── learning_rate.png
│   └── screenshots/        # Manual screenshots
│
└── mri_images/             # Dataset (gitignored, auto-downloaded)
    ├── train/
    └── test/

Model Architecture

graph TB
    subgraph DataPrep["1. Data Preparation"]
        A["Input: Alzheimer MRI Scans<br/>11,519 images 128x128px"] --> B["Train-Test Split<br/>80% Train / 20% Test"]
        B --> C["Image Normalization<br/>Rescaling 1/255"]
    end

    subgraph ModelArch["2. Model Architecture"]
        D["CNN Definition<br/>Keras Sequential"]

        subgraph Layers["Network Layers"]
            E["Conv2D Block 1<br/>64 filters, 3x3, ReLU + MaxPool"] --> F["Conv2D Block 2<br/>32 filters, 3x3, ReLU + MaxPool"]
            F --> G["Conv2D Block 3<br/>32 filters, 2x2, ReLU + MaxPool"]
            G --> H["Dropout<br/>0.20 / 0.25"]
            H --> I["Flatten<br/>8192"]
            I --> J["Dense 100 + Dense 50<br/>Softmax 4 classes"]
        end

        D --> E
    end

    subgraph Training["3. Training Pipeline"]
        K["Model Compilation<br/>Adam Optimizer<br/>Categorical Crossentropy"] --> L["Model Training<br/>94 Epochs (EarlyStopping)<br/>ReduceLROnPlateau"]
    end

    subgraph Evaluation["4. Evaluation and Output"]
        M["Model Evaluation"]

        subgraph Metrics["Performance Metrics"]
            N["Accuracy: 98.83%<br/>Loss Plots"]
            O["Classification Report<br/>Precision, Recall, F1"]
            P["Confusion Matrix"]
        end

        M --> N
        M --> O
        M --> P

        Q["Output: Saved Model<br/>.keras file<br/>Inference: ~2.4ms/image"]
    end

    DataPrep --> ModelArch
    ModelArch --> Training
    Training --> Evaluation
    Training -. "Backpropagation" .-> ModelArch

Loading

Setup

git clone https://github.com/Bugaddr/alzheimer_detection.git
cd alzheimer_detection
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

The dataset is auto-downloaded from Kaggle on first run.

Usage

# Train the model
python train.py

# Evaluate on test set
python test.py

# Generate paper figures (11 figures from real data)
python generate_figures.py

# Run web app
python app.py
# Open http://localhost:9000

Web Interface

The Flask web app lets clinicians upload MRI scans and get instant predictions with confidence scores, clinical recommendations, and a patient-friendly chatbot that explains results in simple language.

API Endpoints:

Method	Endpoint	Description
GET	/	Frontend UI
GET	/api/health	Health check
POST	/api/predict	Analyze an MRI image
POST	/api/chat	Chatbot — interprets results for patients

Chatbot

After a scan result is displayed, a floating chat widget appears. Patients can ask questions like:

• "What does my result mean?"
• "What should I do next?"
• "What lifestyle changes can help?"
• "How accurate is this test?"
• "Should I be worried?"

The chatbot uses a rule-based knowledge system (no external API) tailored to each of the 4 diagnosis levels.

⚠️ Disclaimer: This is a research prototype and clinical decision-support tool, not a standalone diagnostic instrument. All results must be validated by qualified healthcare professionals.

License

MIT