Joint super-resolution and synthesis of 1 mm isotropic MP-RAGE volumes from clinical MRI exams with scans of different orientation, resolution and contrast (2012.13340v1)

Abstract: Most existing algorithms for automatic 3D morphometry of human brain MRI scans are designed for data with near-isotropic voxels at approximately 1 mm resolution, and frequently have contrast constraints as well - typically requiring T1 scans (e.g., MP-RAGE). This limitation prevents the analysis of millions of MRI scans acquired with large inter-slice spacing ("thick slice") in clinical settings every year. The inability to quantitatively analyze these scans hinders the adoption of quantitative neuroimaging in healthcare, and precludes research studies that could attain huge sample sizes and hence greatly improve our understanding of the human brain. Recent advances in CNNs are producing outstanding results in super-resolution and contrast synthesis of MRI. However, these approaches are very sensitive to the contrast, resolution and orientation of the input images, and thus do not generalize to diverse clinical acquisition protocols - even within sites. Here we present SynthSR, a method to train a CNN that receives one or more thick-slice scans with different contrast, resolution and orientation, and produces an isotropic scan of canonical contrast (typically a 1 mm MP-RAGE). The presented method does not require any preprocessing, e.g., skull stripping or bias field correction. Crucially, SynthSR trains on synthetic input images generated from 3D segmentations, and can thus be used to train CNNs for any combination of contrasts, resolutions and orientations without high-resolution training data. We test the images generated with SynthSR in an array of common downstream analyses, and show that they can be reliably used for subcortical segmentation and volumetry, image registration (e.g., for tensor-based morphometry), and, if some image quality requirements are met, even cortical thickness morphometry. The source code is publicly available at github.com/BBillot/SynthSR.

• The paper presents SynthSR, a CNN framework that jointly performs super-resolution and contrast synthesis to produce high-quality isotropic MP-RAGE volumes from clinical MRI scans.
• The methodology leverages synthetic training data and simulates real-world acquisition artifacts to overcome limitations of anisotropic resolution and variable MR contrasts.
• The results show enhanced morphometric analysis with reduced volumetric errors, paving the way for improved diagnostics in conditions like Alzheimer’s disease.

Super-Resolution and Synthesis of Clinical MRI Scans with SynthSR: An Overview

The paper "Joint super-resolution and synthesis of 1 mm isotropic MP-RAGE volumes from clinical MRI exams with scans of different orientation, resolution, and contrast" introduces SynthSR, an advanced method leveraging convolutional neural networks (CNNs) to transform conventional clinical MRI scans into isotropic volumes of standard MR contrast, such as MP-RAGE, without high-resolution training data for the input modalities. This approach is significant due to the vast number of clinical MRI scans conducted worldwide with protocols considerably less comprehensive than those typically utilized in research settings.

Technical Challenges and Motivation

The predominant limitation of existing neuroimaging tools is their dependency on near-isotropic voxel resolutions, roughly 1 mm, and specific MR contrasts like T1-weighted images. This dependency considerably restricts the use of conventional clinical images, often characterized by anisotropic resolutions, larger slice spacing, and varying contrasts, in automated quantitative analysis. Moreover, the acquisition protocols usually differ across clinical sites, posing additional hurdles in morphometric analyses and research applications.

Recent advancements in CNNs have shown promise in super-resolution and contrast synthesis but are bound to datasets resembling their training data's resolution and contrast distributions. SynthSR attempts to fill this gap by creatively utilizing synthetic images for training, enabling CNN models to generalize across diverse clinical MRI acquisition protocols.

Methodology Highlights

SynthSR introduces a generative model that fabricates multi-modal MRI scans capturing variations in resolution and MR contrast. The core of the SynthSR framework lies in generating training datasets from segmentation templates, which undergo simulated processes mimicking real-world acquisition artifacts like bias fields, partial voluming, and registration errors. This synthetic training approach empowers CNNs to execute joint super-resolution and contrast synthesis effectively.

Results and Implications

Through meticulous experimentations, SynthSR illustrates substantial capabilities in increasing resolution and synthesizing isotropic MR volumes, enabling improved automated downstream analyses. Quantitative metrics demonstrate robust outcomes, like reduced hippocampal volumetric errors and superior detection of Alzheimer's disease-related effects using clinical data.

SynthSR bridges the technical void between clinical and research-grade MRI practices, indicating its potential as a tool for amplifying the applicability of vast MRI data repositories stored in clinical archives.

Future Prospects

The methodological innovations herein open pathways for enhancing additional imaging features, optimizing architectural designs, and simulation models involved in the generative procedure. Continuing development might focus on accommodating more complex pathological alterations beyond synthetic artifacts, ensuring SynthSR's efficacy is broadened and even more practical for a myriad of neuroimaging tasks.

SynthSR not only represents an incremental step towards synthesizing high-fidelity images from sparse data but also underscores the paradigm's transformative potential in utilizing pre-existing clinical datasets for extensive neuroimaging research, pushing the boundaries of statistical power and clinical applications in understanding brain morphology and pathology. As neuroimaging communities seek more inclusive approaches, SynthSR provides a model framework.