Synthesis of High-Resolution Research-Quality MRI Data from Clinical MRI Data in Patients with COVID-19

Ryan J. Cali, Holly J. Freeman, Benjamin Billot, Megan E. Barra, David Fischer, William R. Sanders, Susie Y. Huang, John Conklin, Bruce Fischl, Juan Eugenio Iglesias, Brian L. EdlowmedRxiv 2021 · DOI

Abstract

Pathophysiological mechanisms of neurological disorders in patients with coronavirus disease 2019 (COVID-19) are poorly understood, partly because of a lack of high-resolution neuroimaging data. We applied SynthSR, a convolutional neural network that synthesizes high-resolution isotropic research-quality data from thick-slice clinical MRI data, to a cohort of 11 patients with severe COVID-19. SynthSR successfully synthesized T1-weighted MPRAGE data at 1 mm spatial resolution for all 11 patients, each of whom had at least one brain lesion. Correlations between volumetric measures derived from synthesized and acquired MPRAGE data were strong for the cortical grey matter, subcortical grey matter, brainstem, hippocampus, and hemispheric white matter ($r = 0.84$ to $0.96$, $p \leq 0.001$), but absent for the cerebellar white matter and corpus callosum ($r = 0.04$ to $0.17$, $p > 0.61$). SynthSR creates an opportunity to quantitatively study clinical MRI scans and elucidate the pathophysiology of neurological disorders in patients with COVID-19, including those with focal lesions.

Introduction

Up to 80% of hospitalized COVID-19 patients experience neurological symptoms, yet our understanding of the underlying pathophysiology remains limited. A primary bottleneck in this research is data quality. Research-grade MRI requires patients to remain stationary for extended periods to acquire "isotropic" data, where voxels are uniform 1mm cubes.

In clinical settings—particularly for critically ill COVID-19 patients—this is often impossible. To minimize scan time and reduce the risk of viral exposure for staff, clinicians typically acquire "thick-slice" scans with 5–6mm spacing. While these scans are sufficient for a radiologist to identify major strokes or hemorrhages, they are incompatible with automated software used to measure brain volume or cortical thickness. This study demonstrates a method for "upcycling" these standard clinical brain scans into high-resolution data suitable for advanced neuroimaging research.

Methodology: The SynthSR Approach

The researchers employed SynthSR, a joint super-resolution and synthesis method based on a Convolutional Neural Network (CNN). Unlike many deep learning models that require paired high-resolution and low-resolution data for training, SynthSR was trained on highly diverse synthetic images generated from segmentations. This makes the model remarkably robust to variations in scanners and image contrasts.

The Workflow

  1. Input: Two standard clinical sequences (typically T2 FLAIR and T1 or T2) with large inter-slice spacing.
  2. Processing: The sequences are co-registered and fed into a 2-channel CNN.
  3. Output: A synthesized 1mm isotropic T1-weighted MPRAGE image.
  4. Validation: The researchers used FreeSurfer, a standard neuroimaging suite, to segment both the synthesized images and actual "gold-standard" high-resolution images acquired from the same patients to measure accuracy.

Dataset and Clinical Context

The study focused on a retrospective cohort of 11 patients with severe COVID-19 admitted to the Neurosciences ICU at Massachusetts General Hospital. All patients required mechanical ventilation and had a positive PCR test.

The dataset was particularly challenging due to the presence of significant pathology: nine patients had ischemic strokes, and several presented with intracranial hemorrhages or microhemorrhages. This provided a rigorous test for the AI’s ability to handle "out-of-distribution" features like focal brain lesions.

Key Results and Performance

The quantitative validation demonstrated that synthesized volumes were highly reliable for most major brain regions. The correlations between synthesized and acquired data were particularly strong in areas critical for neurodegenerative research:

  • Subcortical Grey Matter: $r = 0.96$
  • Putamen: $r = 0.96$
  • Brainstem: $r = 0.93$
  • Cortical Grey Matter: $r = 0.91$
  • Hippocampus: $r = 0.85$

However, the model reached its limits in regions with complex or thin white matter architecture. It showed almost no correlation for the cerebellar white matter ($r = 0.04$) and the corpus callosum ($r = 0.17$).

Discussion: Robustness and Risks

One of the most significant findings is SynthSR’s robustness to focal lesions. AI models often struggle when encountering anatomy that deviates from the norm (such as a stroke-induced void), yet SynthSR successfully synthesized anatomy around these regions in most cases.

However, the study also identified a "hallucination" risk. In one instance, the AI attempted to "fill in" an area of encephalomalacia (brain damage) with inaccurate synthetic anatomy. This underscores the necessity for human-in-the-loop quality assessment when using AI-synthesized data for clinical or research conclusions.

Conclusion

SynthSR provides a powerful tool for retrospective research, allowing scientists to utilize existing clinical databases to study the neurological impact of COVID-19. By converting low-resolution clinical scans into research-quality data, we can now perform large-scale studies on COVID-19-related atrophy and neurodegeneration, potentially even looking back at early scans of patients who now suffer from "long-COVID" to identify predictive volumetric changes.

The tool has been made available to the community via GitHub and as part of the FreeSurfer software suite.

Summary updated

Abstract

Abstract Pathophysiological mechanisms of neurological disorders in patients with coronavirus disease 2019 (COVID-19) are poorly understood, partly because of a lack of high-resolution neuroimaging data. We applied SynthSR, a convolutional neural network that synthesizes high-resolution isotropic research-quality data from thick-slice clinical MRI data, to a cohort of 11 patients with severe COVID-19. SynthSR successfully synthesized T1-weighted MPRAGE data at 1 mm spatial resolution for all 11 patients, each of whom had at least one brain lesion. Correlations between volumetric measures derived from synthesized and acquired MPRAGE data were strong for the cortical grey matter, subcortical grey matter, brainstem, hippocampus, and hemispheric white matter ( r =0.84 to 0.96, p≤0.001), but absent for the cerebellar white matter and corpus callosum ( r =0.04 to 0.17, p>0.61). SynthSR creates an opportunity to quantitatively study clinical MRI scans and elucidate the pathophysiology of neurological disorders in patients with COVID-19, including those with focal lesions.
BibTeX
@misc{cali2021synthesis,
  author = {Ryan J. Cali and Holly J. Freeman and Benjamin Billot and Megan E. Barra and David Fischer and William R. Sanders and Susie Y. Huang and John Conklin and Bruce Fischl and Juan Eugenio Iglesias and Brian L. Edlow},
  title = {{Synthesis of High-Resolution Research-Quality MRI Data from Clinical MRI Data in Patients with COVID-19}},
  year = {2021},
  doi = {10.1101/2021.11.25.21266090},
  howpublished = {medRxiv},
}
Figure