Abstract
We present a deep network interpolation strategy for accelerated parallel MR image reconstruction. In particular, we examine the network interpolation in parameter space between a source model that is formulated in an unrolled scheme with L1 and SSIM losses and its counterpart that is trained with an adversarial loss. We show that by interpolating between the two different models of the same network structure, the new interpolated network can model a trade-off between perceptual quality and fidelity.
Introduction
Accelerating MRI scans through undersampling—collecting less k-space data—is essential for reducing scan times but often results in aliased or blurry images. Deep learning has emerged as a powerful tool for high-quality reconstruction. However, a persistent trade-off exists between two training paradigms:
- Fidelity-focused models: Trained with standard losses like $L_1$ or SSIM, these models achieve high quantitative scores but often produce over-smoothed images that lack fine clinical textures.
- Perceptual-focused models: Trained with Generative Adversarial Networks (GANs), these models produce sharp, realistic textures but risk "hallucinating" features or introducing artifacts not present in the original data.
This work addresses this challenge by proposing a method to balance these extremes, allowing for a smooth trade-off between image fidelity and perceptual sharpness through network interpolation.
Methods
The proposed strategy utilizes Deep Network Interpolation in the parameter space. The core architecture is a "Sensitivity Network" (SN), an unrolled scheme mimicking iterative optimization. Two versions of this network are trained:
- Model A (SN): Optimized for fidelity using $L_1$ and SSIM losses.
- Model B (SN-GAN): Optimized for perceptual quality using an adversarial loss (LSGAN).
Instead of interpolating the output images, the weights ($\theta$) of the networks are interpolated: $$\theta_{interp} = (1 - \alpha)\theta_{SN} + \alpha\theta_{SN-GAN}$$ By varying the coefficient $\alpha \in [0, 1]$, the network's behavior transitions from conservative fidelity to creative perceptual sharpness without the need for retraining.
Dataset and Experimental Setup
The study utilized the fastMRI knee dataset, comprising 973 training and 199 validation volumes of multi-coil (15-channel) raw k-space data. Evaluations were performed on Proton-Density (PD) and PD-Fat-Suppressed (PDFS) scans at acceleration factors (AF) of 4x and 8x.
Results and Discussion
Quantitative Performance
Quantitative analysis (Table 1) demonstrates that the interpolated model (interpSN) consistently outperforms the standard Unet baseline across NMSE, PSNR, and SSIM metrics. While the $L_1$-optimized SN model typically achieves the highest PSNR (as it is directly optimized for pixel-wise fidelity), the interpolated model maintains competitive scores while offering significantly improved visual clarity.
Qualitative Insights
Interpolating in the parameter space ensures structural consistency while shifting texture properties. The transition is continuous and smooth, as illustrated in Figure 1. This "zero-cost tuning" allows for real-time adjustment of image sharpness, potentially enabling radiologists to find an optimal balance for specific diagnostic tasks. The interpolated model at $\alpha=0.5$ often provides a superior balance, recovering fine details blurred by standard $L_1$ models without the potential artifacts introduced by pure GAN reconstructions.
Conclusion
Deep network interpolation provides a flexible framework for MRI reconstruction, bridging the gap between quantitative fidelity and perceptual quality. This approach allows for the generation of high-quality images that are both quantitatively accurate and visually sharp. Future work will focus on automated selection of the interpolation coefficient and formal clinical validation with radiologists to assess the diagnostic impact of enhanced texture sharpness.
Summary updated