RIRGAN: An end-to-end lightweight multi-task learning method for brain MRI super-resolution and denoising

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Single-task learning (STL) models in medical imaging focus on either super-resolution (SR) or denoising (DN).
• Real-world medical images often suffer from both low resolution and high noise, limiting diagnostic accuracy.
• Existing methods are insufficient for addressing these dual challenges simultaneously.

Purpose of the Study:

• To develop an efficient model capable of performing both super-resolution and denoising tasks concurrently for low-level vision medical images.
• To address the limitations of single-task learning approaches in practical medical imaging scenarios.
• To improve the quality and diagnostic utility of medical images affected by low resolution and high noise.

Main Methods:

• Proposed an end-to-end lightweight multi-task learning (MTL) generative adversarial network (GAN) named RIRGAN.
• Employed residual-in-residual blocks (RIR-Blocks) for feature extraction and a long skip connection (LSC) for deep network construction.
• Integrated a relativistic average discriminator (RaD) and a hybrid loss function for enhanced image generation and balanced performance.

Main Results:

• RIRGAN demonstrated superior performance in both subjective perception and objective evaluation metrics compared to STL-based SR and DN methods.
• The model effectively handles medical images with simultaneous low resolution and high noise.
• Generated images exhibit more realistic details and improved high-frequency information preservation.

Conclusions:

• RIRGAN offers a robust solution for enhancing low-level vision medical images by performing super-resolution and denoising simultaneously.
• The proposed MTL approach aligns better with the practical requirements of medical practice.
• This work advances deep learning applications in medical image analysis, improving diagnostic capabilities.