Invertible Sharpening Network for MRI Reconstruction Enhancement

Siyuan Dong, Eric Z. Chen, Lin Zhao, Xiao Chen, Yikang Liu, Terrence Chen, Shanhui Sun

Abstract

High-quality MRI reconstruction is essential for clinical diagnostics, yet many deep learning methods produce blurry results by optimizing for standard metrics like PSNR and SSIM. To address this, we introduce InvSharpNet, an invertible sharpening network designed to enhance visual quality. Unlike traditional approaches that attempt to map blurry inputs to sharp ground truths, InvSharpNet employs a backward training strategy: it learns a blurring transform from the ground truth to the reconstruction. Due to the network's mathematical invertibility, this process can be reversed during inference to sharpen images. Experiments across multiple MRI datasets and evaluations by radiologists demonstrate that InvSharpNet significantly improves sharpness and diagnostic confidence with minimal artifacts.

Introduction

Magnetic Resonance Imaging (MRI) is a cornerstone of modern medicine, but its slow acquisition speed often necessitates under-sampling. While deep learning models can reconstruct full images from these sparse measurements, they frequently suffer from the "solution averaging" problem. Because reconstruction is an ill-posed "one-to-many" problem, standard training using Mean Squared Error (MSE) often results in a pixel-wise average of all possible sharp solutions, leading to a safe but blurry output. While Generative Adversarial Networks (GANs) can produce sharper images, they are prone to "hallucinating" details that do not exist, which poses a significant risk in medical settings.

Methodology

The Backward Training Strategy

The core innovation of InvSharpNet is its reversal of the standard training paradigm. Instead of forcing a network to solve the complex task of recovering lost high-frequency details from a blurry image, the authors train the network to perform the well-defined, one-to-one task of mapping a sharp ground-truth image to its blurry reconstruction.

Invertible Architecture

InvSharpNet utilizes an Invertible Neural Network (INN) architecture based on iResNet. Because the mapping from "sharp to blurry" is learned using an invertible structure, the network can be mathematically flipped during deployment. During inference, the blurry reconstruction is fed into the network in the forward direction to produce a sharpened output.

To ensure this invertibility, the network must satisfy a Lipschitz constraint. The authors found a "sweet spot" with a Lipschitz constant of 0.7; this value provides enough learning capacity to capture complex image features while maintaining the stability required for fixed-point iterations during the inversion process.

Experimental Setup

The model was rigorously tested on three distinct datasets to ensure robustness and generalizability:

  • fastMRI Knee dataset: Comprising over 34,000 training images.
  • fastMRI Brain dataset: Comprising over 67,000 training images.
  • In-house Brain dataset: Featuring five different gradient echo (GRE) contrasts to test the model's performance on data outside the fastMRI ecosystem.

Results and Discussion

Perceptual Quality and Metrics

The study highlights a common pitfall in medical imaging research: standard metrics like PSNR and SSIM often penalize sharp images because they prioritize pixel-wise similarity over perceptual clarity. InvSharpNet excelled in LPIPS (Learned Perceptual Image Patch Similarity), which more closely aligns with human vision.

Clinical Validation

A blind study involving two professional radiologists confirmed the clinical utility of the approach. InvSharpNet (backward) received a sharpness rating of 4.67/5.0, compared to 4.00/5.0 for the baseline reconstruction. Crucially, InvSharpNet maintained high artifact-resistance scores (4.00/5.0), avoiding the hallucinated features typically seen in GAN-based methods (e.g., cWGAN).

Generalizability

The model demonstrated strong performance on the in-house GRE dataset despite being trained on fastMRI data. This suggests that the sharpening transform learned by InvSharpNet is robust across different scanner types and pulse sequences, making it a viable candidate for diverse clinical environments.

Conclusion and Future Work

InvSharpNet offers a novel solution to the blurriness inherent in deep learning-based MRI reconstructions. By leveraging backward training and invertible architectures, it provides a mathematically grounded way to recover sharpness without the risks of hallucination. While the current implementation requires more computational resources and a larger parameter count (5.4M) to maintain invertibility, the results represent a significant step forward in perceptual MRI quality. Future research may extend this backward training approach to other image restoration tasks, such as denoising and super-resolution.

Summary updated

Abstract

High-quality MRI reconstruction plays a critical role in clinical applications. Deep learning-based methods have achieved promising results on MRI reconstruction. However, most state-of-the-art methods were designed to optimize the evaluation metrics commonly used for natural images, such as PSNR and SSIM, whereas the visual quality is not primarily pursued. Compared to the fully-sampled images, the reconstructed images are often blurry, where high-frequency features might not be sharp enough for confident clinical diagnosis. To this end, we propose an invertible sharpening network (InvSharpNet) to improve the visual quality of MRI reconstructions. During training, unlike the traditional methods that learn to map the input data to the ground truth, InvSharpNet adapts a backward training strategy that learns a blurring transform from the ground truth (fully-sampled image) to the input data (blurry reconstruction). During inference, the learned blurring transform can be inverted to a sharpening transform leveraging the network's invertibility. The experiments on various MRI datasets demonstrate that InvSharpNet can improve reconstruction sharpness with few artifacts. The results were also evaluated by radiologists, indicating better visual quality and diagnostic confidence of our proposed method.
BibTeX
@inproceedings{dong2022invertible,
  author = {Siyuan Dong and Eric Z. Chen and Lin Zhao and Xiao Chen and Yikang Liu and Terrence Chen and Shanhui Sun},
  title = {{Invertible Sharpening Network for MRI Reconstruction Enhancement}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2022},
  url = {https://openreview.net/forum?id=y4Gv8I5L9Y},
}
Figure