HADESR-GAN: A Lightweight Hybrid Attention-based GAN for Medical Image Super-Resolution

This repository contains the official PyTorch implementation of HADESR-GAN, a parameter-efficient hybrid-attention generative adversarial network for medical image super-resolution, as presented in the paper "HADESR-GAN: A Lightweight Hybrid Attention-based GAN for Medical Image Super-Resolution." which is already submitted to Nature Scientific Reports

---

🧠 Abstract

High-resolution (HR) medical imaging is vital for accurate diagnosis, yet its deployment is often constrained by hardware limitations, radiation exposure, and real-time processing demands.
We introduce HADESR-GAN, a parameter-efficient hybrid-attention generative adversarial network for medical image super-resolution.
The generator integrates the Residual Whole Map Attention (RWMAN) backbone of MedSRGAN with a lightweight Hybrid Attention Separable Block (HASB) based on depthwise-separable convolutions and attention fusion.
Transformer-inspired global context modeling further enhances textural consistency.
This model achieves high performance with approximately 85% fewer parameters than its baseline, MedSRGAN.

---

✨ Key Features

• Lightweight Architecture: Achieves state-of-the-art results with only ~1.6 million parameters.
• Hybrid Attention Mechanism: Integrates a Hybrid Attention Separable Block (HASB) with a Residual Whole Map Attention (RWMAN) backbone.
• High Performance: Achieves excellent PSNR and SSIM scores across multiple medical imaging datasets.
• GAN-based Training: Produces perceptually realistic high-resolution images.
• Modular and Configurable: The code is organized for easy modification and experimentation.

---

🏗️ Model Architecture

The HADESR-GAN consists of a Generator and a Discriminator, trained in an adversarial setup.

🔹 Generator

The generator combines Short Residual Connections (SRC) with the novel Hybrid Attention Separable Blocks (HASB), significantly reducing the number of layers compared to the original MedSRGAN.

🔹 Discriminator

The discriminator is a deep CNN with a dual-path design for multi-scale analysis, enabling it to effectively distinguish between real and super-resolved images.

---

📊 Results

HADESR-GAN was evaluated on three medical imaging datasets:

Dataset	Domain	PSNR (dB)	SSIM
Chest CT (LIDC-IDRI)	CT	34.32	0.90
Echocardiography (Tehran Heart Center)	Echo	32.04	0.92
Dermoscopy (ISIC)	Skin	37.18	0.94

📁 Project Structure

HARNSRGAN/
│
├── dataset/
│ ├── train/
│ └── val/
│
├── models/
│ ├── discriminator.py
│ └── generator.py
│
├── losses/
│ ├── discriminator_loss.py
│ └── generator_loss.py
│
├── trainer/
│ └── trainer.py
│
├── utils/
│ ├── dataset.py
│ ├── early_stopping.py
│ └── visualize.py
│
├── config.py
├── main.py
├── requirements.txt
└── README.md

---

⚙️ Setup and Usage

🧩 Prerequisites

• Python 3.8+
• PyTorch
• Other dependencies listed in requirements.txt

💻 Installation

Clone the repository:

git clone https://github.com/parsakhavarinejad/HADESR-GAN.git
cd HADESR-GAN

Install dependencies:

pip install -r requirements.txt

🚀 Training

Prepare your dataset: Place training images in dataset/train/ and validation images in dataset/val/.

Configure training parameters: Modify config.py as needed.

Run the training script:

python main.py

🧾 Citation

If you find this work useful for your research, please cite our paper when it is published.