Coupled-Projection Residual Network for MRI Super-Resolution

Chun-Mei Feng, Kai Wang, Shijian Lu, Yong Xu, Heng Kong, Ling ShaoarXiv 2019 · OpenReview

Abstract

Magnetic Resonance Imaging (MRI) is a critical tool in clinical diagnosis and pathology research. However, obtaining high-resolution (HR) images remains a challenge due to acquisition time and hardware constraints. This paper introduces the Coupled-Projection Residual Network (CPRN), a novel architecture for MRI super-resolution. The CPRN utilizes two complementary sub-networks: a shallow network that ensures structural content consistency through a feedback-driven coupled-projection mechanism, and a deep network that recovers high-frequency residuals using cascaded residual blocks. To optimize feature fusion, a step-wise connection (CPRN_S) is introduced, inspired by human cognitive processes. Experimental results across three public MRI datasets (Brats and ATLAS) demonstrate that CPRN achieves superior performance in both quantitative metrics (PSNR/SSIM) and qualitative visual fidelity compared to state-of-the-art methods.

Introduction

MRI is essential for accurate medical diagnosis, yet high-resolution scans are often slow to acquire and sensitive to patient motion. Super-resolution (SR) techniques offer a computational path to enhance low-resolution (LR) scans. Traditional deep learning SR models often struggle with either excessive computational complexity or a failure to recover fine, medically relevant textures. The Coupled-Projection Residual Network (CPRN) addresses these issues by separating the task into structural preservation and fine-detail recovery, linked by an efficient fusion strategy.

Methodology

The CPRN architecture is divided into two primary pathways designed to handle different aspects of the image reconstruction process.

Dual-Pathway Architecture

  1. Shallow Sub-network (Coupled-Projection): This component focuses on maintaining "content consistency." It employs a feedback mechanism involving iterative up-projection and down-projection. By calculating the projection error at each step, the network self-corrects to ensure the reconstructed high-resolution features align perfectly with the original structural layout of the LR input.
  2. Deep Sub-network (Residual Learning): While the shallow network handles global structure, the deep network focuses on "residuals"—the high-frequency differences, such as fine edges and textures, between LR and HR images. It utilizes multiple cascaded residual blocks to learn these complex patterns without the constraints of Batch Normalization, which the authors found can restrict feature flexibility in medical imaging.
The CPRN architecture consists of a shallow network (Part A) for structural consistency and a deep network (Part B) for residual learning.

Step-wise Connection (CPRN_S)

Inspired by the human cognitive process of learning from simple to complex, the authors developed a "step-wise" variant (CPRN_S). This design fuses features from the shallow and deep layers progressively rather than all at once. This approach facilitates better feature reuse and allows the model to maintain high performance while reducing the total parameter count by approximately 30%.

Experimental Setup

The model was rigorously evaluated on three public brain MRI datasets:

  • Brats: Preoperative multimodal images of glioblastoma and lower-grade gliomas.
  • ATLAS (native & standardized): Anatomical Tracings of Lesions After Stroke, focusing on stroke-related brain pathology.

Training involved down-sampling T1-weighted HR images via bicubic interpolation to create LR/HR pairs for supervised learning.

Results and Discussion

Quantitative Performance

CPRN consistently outperformed established models including SRCNN, VDSR, EDSR, and DDBPN. On the Brats dataset at a x2 scale, CPRN achieved a Peak Signal-to-Noise Ratio (PSNR) of 36.500, surpassing the significantly heavier EDSR (36.123) and DDBPN (36.158).

Qualitative Visual Fidelity

Visual inspection confirms that CPRN excels at recovering sharp boundaries and fine textures. While other models often produce blurry edges or artificial "stripe patterns," CPRN successfully reconstructs clear blood vessel boundaries and intricate gray matter textures, which are vital for clinical assessment.

Qualitative comparison of CPRN against state-of-the-art methods across Brats and ATLAS datasets, showing superior texture recovery and higher PSNR/SSIM values.

Efficiency and Complexity

The CPRN_S variant represents a significant advancement in balancing accuracy and efficiency. By utilizing the step-wise connection, the model achieves competitive results with a leaner architecture, making it more viable for deployment in clinical environments where computational resources may be limited.

Conclusion

The Coupled-Projection Residual Network (CPRN) provides a robust solution for MRI super-resolution by effectively decoupling structural preservation from texture recovery. The introduction of the feedback-driven shallow network and the cognitive-inspired step-wise connection allows for high-fidelity reconstructions that surpass current state-of-the-art methods. Future work will explore the application of this architecture to other MRI modalities, such as cardiac and spinal imaging, to further validate its versatility in medical diagnostics.

Summary updated

Abstract

Magnetic Resonance Imaging(MRI) has been widely used in clinical application and pathology research by helping doctors make more accurate diagnoses. On the other hand, accurate diagnosis by MRI remains a great challenge as images obtained via present MRI techniques usually have low resolutions. Improving MRI image quality and resolution thus becomes a critically important task. This paper presents an innovative Coupled-Projection Residual Network (CPRN) for MRI super-resolution. The CPRN consists of two complementary sub-networks: a shallow network and a deep network that keep the content consistency while learning high frequency differences between low-resolution and high-resolution images. The shallow sub-network employs coupled-projection for better retaining the MRI image details, where a novel feedback mechanism is introduced to guide the reconstruction of high-resolution images. The deep sub-network learns from the residuals of the high-frequency image information, where multiple residual blocks are cascaded to magnify the MRI images at the last network layer. Finally, the features from the shallow and deep sub-networks are fused for the reconstruction of high-resolution MRI images. For effective fusion of features from the deep and shallow sub-networks, a step-wise connection (CPRN S) is designed as inspired by the human cognitive processes (from simple to complex). Experiments over three public MRI datasets show that our proposed CPRN achieves superior MRI super-resolution performance as compared with the state-of-the-art. Our source code will be publicly available at http://www.yongxu.org/lunwen.html.
BibTeX
@misc{feng2019coupled,
  author = {Chun-Mei Feng and Kai Wang and Shijian Lu and Yong Xu and Heng Kong and Ling Shao},
  title = {{Coupled-Projection Residual Network for MRI Super-Resolution}},
  year = {2019},
  eprint = {1907.05598},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1907.05598},
}
Figure