[Paper Review] A Tool for Super-Resolving Multimodal Clinical MRI

Mikael Brudfors, Yaël Balbastre+2 more|arXiv (Cornell University)|Sep 3, 2019

Generative Adversarial Networks and Image Synthesis49 references4 citations

TL;DR

This paper presents a data-agnostic, model-based super-resolution tool for multimodal clinical MRI that reconstructs high-resolution isotropic images from low-resolution, thick-sliced scans using a generative model with multi-channel total variation regularization. The method achieves superior performance over conventional techniques on simulated and real clinical data without requiring training or manual tuning.

ABSTRACT

We present a tool for resolution recovery in multimodal clinical magnetic resonance imaging (MRI). Such images exhibit great variability, both biological and instrumental. This variability makes automated processing with neuroimaging analysis software very challenging. This leaves intelligence extractable only from large-scale analyses of clinical data untapped, and impedes the introduction of automated predictive systems in clinical care. The tool presented in this paper enables such processing, via inference in a generative model of thick-sliced, multi-contrast MR scans. All model parameters are estimated from the observed data, without the need for manual tuning. The model-driven nature of the approach means that no type of training is needed for applicability to the diversity of MR contrasts present in a clinical context. We show on simulated data that the proposed approach outperforms conventional model-based techniques, and on a large hospital dataset of multimodal MRIs that the tool can successfully super-resolve very thick-sliced images. The implementation is available from https://github.com/brudfors/spm_superres.

Motivation & Objective

• To address the challenge of automated neuroimaging analysis on highly variable clinical MRI scans with thick slices and diverse contrasts.
• To develop a data-agnostic method that does not require retraining for new scan contrasts or resolution factors.
• To enable high-fidelity super-resolution of clinical MRI by fusing information across multiple MR contrasts.
• To overcome limitations of interpolation-based upsampling, which introduces artifacts and biases downstream analysis.
• To provide a robust, generalizable tool for large-scale clinical studies using existing, population-representative MRI data.

Proposed method

• The method employs a generative model that infers a high-resolution (HR) image from low-resolution (LR) multimodal MRI scans using a joint likelihood formulation.
• It uses a multi-channel total variation (MTV) regularization to exploit gradient information across different MRI contrasts, enhancing edge preservation.
• The forward model incorporates slice profile and gap parameters, with fixed values assumed for simplicity, though extension to estimate them is discussed.
• A major component is the use of an ADMM algorithm to solve the nonsmooth optimization problem arising from the TV prior.
• The model assumes a Gaussian likelihood for the observed LR data and uses a non-stationary penalty via reweighted least-squares to improve convergence and robustness.
• The framework allows for principled marginalization of the HR image and could be extended with variational Bayes or multivariate Gaussian mixtures for joint segmentation and reconstruction.

Experimental results

Research questions

• RQ1 Can a model-based super-resolution approach effectively reconstruct high-resolution isotropic MRI from thick-sliced, multimodal clinical scans without requiring training?
• RQ2 How does the proposed method compare to conventional interpolation and model-based techniques in terms of image quality and artifact suppression?
• RQ3 To what extent does multi-channel total variation improve reconstruction fidelity compared to single-channel regularization?
• RQ4 Can the method generalize across diverse clinical MRI contrasts and scanner parameters without retraining?
• RQ5 What impact does the inclusion of cross-contrast gradient information have on reducing stair-casing artifacts in flat tissue regions?

Key findings

• The proposed method outperforms conventional model-based super-resolution techniques on simulated data, particularly in preserving structural details and reducing artifacts.
• On a large real-world clinical dataset, the tool successfully super-resolved very thick-sliced MRI scans, producing images of quality suitable for downstream automated analysis.
• The use of multi-channel total variation significantly reduces stair-casing effects, as gradients in one contrast can compensate for flat gradients in another.
• The method is robust to variations in slice profile and gap, though performance improves when these parameters are accurately known or estimated.
• The absence of training requirements makes the tool highly generalizable across diverse clinical MRI protocols and scanner types.
• The implementation is publicly available on GitHub, enabling reproducibility and integration into clinical and research pipelines.