Alzheimer's Disease Detection

Overview

Deep learning system for automated Alzheimer’s Disease detection and staging from brain MRI scans using ensemble methods. Classifies patients into 4 categories: Non-Demented, Very Mild Demented, Mild Demented, and Moderate Demented.

💻 GitHub: alzheimers-disease-detection
🧠 Task: Multi-class medical image classification
🎯 Accuracy: 95.6% on test set

Medical Background

Alzheimer’s Disease: Progressive neurodegenerative disorder

• 6.7M Americans living with Alzheimer’s (2023)
• Early detection crucial for treatment planning
• MRI reveals brain atrophy patterns

Clinical Stages:

1. Non-Demented: Normal cognitive function
2. Very Mild: Subtle memory issues (CDR 0.5)
3. Mild Demented: Noticeable impairment (CDR 1)
4. Moderate Demented: Significant cognitive decline (CDR 2)

Model Architecture

Ensemble Approach

Combines predictions from two pre-trained CNNs:

1. EfficientNet-B2

• 9.1M parameters
• Compound scaling (depth + width + resolution)
• Pre-trained on ImageNet

2. VGG16

• 138M parameters
• Deep architecture with small 3×3 filters
• Strong feature extraction

Ensemble Strategy

# Weighted averaging of predictions
ensemble_pred = (0.6 * efficientnet_pred) + (0.4 * vgg16_pred)
final_class = argmax(ensemble_pred)

Rationale: EfficientNet-B2 for efficiency + VGG16 for robustness

Transfer Learning Pipeline

import tensorflow as tf
from tensorflow.keras.applications import EfficientNetB2, VGG16

# EfficientNet-B2 branch
efficientnet = EfficientNetB2(weights='imagenet', include_top=False, input_shape=(224, 224, 3))
efficientnet.trainable = False  # Freeze base layers

x1 = tf.keras.layers.GlobalAveragePooling2D()(efficientnet.output)
x1 = tf.keras.layers.Dense(512, activation='relu')(x1)
x1 = tf.keras.layers.Dropout(0.5)(x1)
output1 = tf.keras.layers.Dense(4, activation='softmax', name='efficientnet_output')(x1)

# VGG16 branch
vgg16 = VGG16(weights='imagenet', include_top=False, input_shape=(224, 224, 3))
vgg16.trainable = False

x2 = tf.keras.layers.GlobalAveragePooling2D()(vgg16.output)
x2 = tf.keras.layers.Dense(512, activation='relu')(x2)
x2 = tf.keras.layers.Dropout(0.5)(x2)
output2 = tf.keras.layers.Dense(4, activation='softmax', name='vgg16_output')(x2)

# Ensemble
ensemble_output = tf.keras.layers.Average()([output1, output2])

Dataset

Source: Alzheimer’s Dataset (4 class of Images)
Samples: 6,400 MRI scans
Split: 80% train, 10% validation, 10% test

Class Distribution:

• Non-Demented: 3,200 images
• Very Mild Demented: 2,240 images
• Mild Demented: 896 images
• Moderate Demented: 64 images

Preprocessing:

• Resize: 224×224
• Normalization: [0, 1] scaling
• Augmentation: rotation (±15°), width/height shift (0.2), horizontal flip

Results

Model Performance

Per-Class Performance (Ensemble)

Confusion Matrix Insights

• High accuracy on Non-Demented and Very Mild stages
• Some confusion between Very Mild ↔ Mild (expected clinically)
• Limited data for Moderate class affects recall

Training Details

Hyperparameters

EPOCHS = 50
BATCH_SIZE = 32
LEARNING_RATE = 0.001
OPTIMIZER = tf.keras.optimizers.Adam(lr=LEARNING_RATE)
LOSS = 'categorical_crossentropy'

Regularization

• Dropout: 0.5 after dense layers
• L2 Regularization: 0.01 for dense layers
• Early Stopping: patience=10, monitor=’val_loss’
• ReduceLROnPlateau: factor=0.5, patience=5

Data Augmentation

from tensorflow.keras.preprocessing.image import ImageDataGenerator

train_datagen = ImageDataGenerator(
    rotation_range=15,
    width_shift_range=0.2,
    height_shift_range=0.2,
    horizontal_flip=True,
    zoom_range=0.2,
    rescale=1./255
)

Tech Stack

Deep Learning: TensorFlow 2.x, Keras
Pre-trained Models: EfficientNet, VGG16, ResNet
Image Processing: OpenCV, Pillow
Visualization: Matplotlib, Seaborn
Deployment: Flask (web interface)

Repository Structure

alzheimers-disease-detection/
├── data/
│   ├── train/
│   ├── val/
│   └── test/
├── models/
│   ├── efficientnet_model.h5
│   ├── vgg16_model.h5
│   └── ensemble_model.h5
├── notebooks/
│   ├── EDA.ipynb
│   ├── model_training.ipynb
│   └── evaluation.ipynb
├── src/
│   ├── preprocess.py
│   ├── train.py
│   ├── evaluate.py
│   └── predict.py
├── app/
│   ├── flask_app.py
│   └── templates/
└── requirements.txt

Clinical Impact

Diagnostic Support

✅ Early Detection: Identify subtle brain changes
✅ Objective Assessment: Quantitative staging
✅ Scalability: Rapid screening for large populations
✅ Consistency: Reduce inter-rater variability

Workflow Integration

1. Radiologist Upload: MRI scan to system
2. Automated Analysis: Ensemble prediction in <5 seconds
3. Probability Report: Confidence scores for each stage
4. Clinical Review: Radiologist validates prediction

Grad-CAM Visualization

Explain Predictions: Where the model looks

import tensorflow as tf
from tf_keras_vis.gradcam import Gradcam

# Generate heatmap
gradcam = Gradcam(model)
cam = gradcam(loss, seed_input, penultimate_layer=-1)

# Overlay on MRI
heatmap = cv2.applyColorMap(cam, cv2.COLORMAP_JET)
superimposed = cv2.addWeighted(original_image, 0.6, heatmap, 0.4, 0)

Findings: Model focuses on hippocampus and ventricles (clinically relevant regions)

Future Enhancements

🔬 3D MRI Analysis: Leverage volumetric scans
🧬 Multimodal Fusion: Combine MRI + PET + CSF biomarkers
📊 Longitudinal Modeling: Track disease progression over time
🏥 Federated Learning: Train on distributed hospital data

Limitations

⚠️ Class Imbalance: Limited Moderate Demented samples
⚠️ Dataset Bias: Single imaging protocol/scanner
⚠️ Generalization: Needs validation on external cohorts
⚠️ Regulatory: Requires FDA/CE approval for clinical use

Status: Academic Project
License: MIT
Contributors: Open to collaboration
Last Updated: 2024