Arga23/dementia-cgan-mri

Dementia Brain MRI Generation using Conditional DCGAN

Model Description

This is a Conditional Deep Convolutional GAN (DCGAN) trained to generate synthetic brain MRI images for different Alzheimer's dementia severity stages.

Approach: noise + stage_label → synthetic_brain_MRI

Model evaluation is conducted qualitatively through visual inspection.

Severity Stages (OASIS Dataset)

• Stage 0: Non-Dementia (Normal)
• Stage 1: Very Mild Dementia
• Stage 2: Mild Dementia
• Stage 3: Moderate Dementia

Quick Start

Option 1: Using inference.py script

# Install dependencies
pip install torch torchvision huggingface_hub pillow

# Download model_architecture.py and inference.py from this repo
# Then run:
python inference.py

The script will automatically:

• Download the model from HuggingFace Hub
• Generate samples for all 4 dementia stages
• Save as generated_stage_0.png, generated_stage_1.png, etc.

Option 2: Python code

import torch
from huggingface_hub import hf_hub_download
from PIL import Image

# Download model architecture first
# (Get model_architecture.py from this repo)
from model_architecture import Generator

# Download generator checkpoint from HuggingFace
model_path = hf_hub_download(
    repo_id="Arga23/dementia-cgan-mri",
    filename="cDCGAN_generator.pth"
)

# Load checkpoint
device = "cuda" if torch.cuda.is_available() else "cpu"
checkpoint = torch.load(model_path, map_location=device)

# Initialize Generator
G = Generator(
    z_dim=checkpoint['z_dim'],
    num_classes=checkpoint['num_classes'],
    img_channels=1
).to(device)

G.load_state_dict(checkpoint['model'])
G.eval()

# Generate image for specific dementia stage
def generate_sample(stage):
    with torch.no_grad():
        z = torch.randn(1, checkpoint['z_dim'], 1, 1).to(device)
        label = torch.tensor([stage]).to(device)
        img = G(z, label)
        img = (img.squeeze().cpu() + 1) / 2
        img = torch.clamp(img, 0, 1)
        return Image.fromarray((img.numpy() * 255).astype('uint8'), mode='L')

# Example: Generate Stage 2 (Mild Dementia)
img = generate_sample(stage=2)
img.save('generated_mild_dementia.png')

Files in Repository

• cDCGAN_generator.pth: Generator weights (inference-ready)
• model_architecture.py: PyTorch Generator architecture
• inference.py: Ready-to-use inference script (downloads from HuggingFace)

Model Architecture

• Input: 100-dim noise vector + stage label (0-3)
• Generator: ConvTranspose2d layers with BatchNorm + ReLU
• Output: 128x128 grayscale MRI image (normalized to [-1, 1])
• Training: Balanced sampling across all 4 stages

Requirements

torch>=2.0.0
torchvision
huggingface_hub
pillow

Disclaimer

⚠️ FOR RESEARCH AND EDUCATIONAL PURPOSES ONLY

Generated images should NOT be used for clinical diagnosis or medical decisions. This model is trained on the OASIS dataset for research purposes only.