Cardiac MRI Image Synthesis using Diffusion Models

This project implements a Denoising Diffusion Probabilistic Model (DDPM) for generating high-resolution cardiac MRI images. The framework leverages attention-enhanced UNet architectures and progressive noise scheduling to reconstruct realistic grayscale images from noise.

Overview

• Architecture: Modified UNet with attention blocks.
• Dataset: Cardiac cine MRI slices in grayscale.
• Diffusion Framework: diffusers library with custom scheduler and training logic.
• Training: Mixed precision with accelerate, cosine learning rate decay, and checkpoint resumption.
• Output: Synthesized images and models saved per defined intervals.

Components

• main.py: Launches training with configuration, dataloaders, and scheduler.
• model.py: Defines the attention-augmented UNet structure.
• train.py: Training loop with logging, evaluation, and checkpointing.
• generate.py: Generates synthetic MRI samples using a trained model.
• dataset.py: Custom PyTorch Dataset with augmentations.
• config.py: Training hyperparameters and device setup.
• utils.py: Utility functions for evaluation, checkpoint loading, and image grid creation.

Usage

python main.py