# Coupled-Projection Residual Network for MRI Super-Resolution

**Authors:** Chun-Mei Feng, Kai Wang, Shijian Lu, Yong Xu, Heng Kong, Ling Shao  
**arXiv:** [1907.05598](https://arxiv.org/abs/1907.05598)  
**Date:** 2019-07-12  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1907.05598v1/

## Key points

- Magnetic Resonance Imaging (MRI) has revolutionized clinical application and pathology research
- Low resolution of MRI images obtained through current techniques is a major challenge
- Introduction of an innovative approach called the Coupled-Projection Residual Network (CPRN) for MRI super-resolution
- CPRN consists of two complementary sub-networks: a shallow network and a deep network
- Shallow sub-network utilizes coupled-projection and a novel feedback mechanism for reconstruction guidance
- Deep sub-network learns from residual high-frequency information using multiple cascaded blocks to magnify MRI images
- Features from both sub-networks are fused together through CPRN S, inspired by human cognitive processes
- Experimental results show that CPRN outperforms state-of-the-art methods in terms of super-resolution performance
- Authors plan to make their source code publicly available for further exploration and utilization by researchers and practitioners

## Summary

, , , , 

Magnetic Resonance Imaging (MRI) has revolutionized clinical application and pathology research by enabling doctors to make more accurate diagnoses. The low resolution of MRI images obtained through current techniques is a major challenge, making it crucial to improve their quality and resolution. This paper introduces an innovative approach called the Coupled-Projection Residual Network (CPRN) for MRI super-resolution. The CPRN consists of two complementary sub-networks: a shallow network and a deep network, which aim to maintain content consistency while learning high-frequency differences between low-resolution and high-resolution images. The shallow sub-network utilizes coupled-projection to better retain image details, incorporating a novel feedback mechanism for reconstruction guidance. Meanwhile, the deep sub-network learns from residual high-frequency information using multiple cascaded blocks at the last layer of the network to magnify MRI images. To effectively reconstruct high-resolution images, features from both sub-networks are fused together through a step-wise connection called CPRN S, inspired by human cognitive processes that progress from simple to complex representations. Experimental results on three public MRI datasets show that CPRN outperforms state-of-the-art methods in terms of super-resolution performance, demonstrating its potential in enhancing MRI image quality and resolution. The authors plan to make their source code publicly available at http://www.yongxu.org/lunwen.html for further exploration and utilization by researchers and practitioners in improving MRI super-resolution techniques.