A Network with Composite Loss and Parameter-free Chunking Fusion Block for Super-Resolution MR Image.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Magnetic Resonance Imaging (MRI) quality is affected by various factors, necessitating high-resolution restoration.
• Single Image Super-Resolution (SISR) using neural networks offers a cost-effective solution for enhancing low-resolution MR images.
• Deep neural networks in SISR can suffer from overfitting and shallow networks struggle with learning complex features.

Purpose of the Study:

• To propose a novel end-to-end super-resolution (SR) method specifically designed for magnetic resonance (MR) images.
• To address the challenges of overfitting and inadequate feature learning in existing deep neural network-based SR methods.
• To improve the accuracy and reliability of MR image super-resolution.

Main Methods:

• Introduced a parameter-free chunking fusion block (PCFB) for improved feature fusion through channel splitting and parameter-free attention.
• Developed a comprehensive training strategy incorporating perceptual loss, gradient loss, and L1 loss to enhance model fitting and prediction accuracy.
• Evaluated the proposed model and training strategy on the IXISR dataset (PD, T1, and T2) for MR image super-resolution.

Main Results:

• The proposed PCFB effectively fuses features by splitting channels, enabling parameter-free attention.
• The combined training strategy significantly improved model fitting accuracy and predictive performance.
• The developed SR method demonstrated advanced performance compared to existing state-of-the-art methods on the IXISR dataset.

Conclusions:

• The novel end-to-end SR method and training strategy effectively enhance MR image resolution.
• The proposed approach overcomes limitations of traditional deep learning methods, achieving superior results in highly reliable MR image measurement.
• This work contributes a robust solution for high-resolution MR image restoration.