Abstract
The blood oxygen level dependent (BOLD) signal from functional magnetic resonance imaging (fMRI) is a noninvasive technique that has been widely used in research to study brain function. However, fMRI suffers from susceptibility-induced off resonance fields which may cause geometric distortions and mismatches with anatomical images. State-of-the-art correction methods require acquiring reverse phase encoded images or additional field maps to enable distortion correction. However, not all imaging protocols include these additional scans and thus cannot take advantage of these susceptibility correction capabilities. As such, in this study we aim to enable state-of-the-art distortion correction with FSL’s topup algorithm of historical and/or limited fMRI data that include only a structural image and single phase encoded fMRI. To do this, we use 3D U-net models to synthesize undistorted fMRI BOLD contrast images from the structural image and use this undistorted synthetic image as an anatomical target for distortion correction with topup. We evaluate the efficacy of this approach, named SynBOLD-DisCo (synthetic BOLD images for distortion correction), and show that BOLD images corrected using our approach are geometrically more similar to structural images than the distorted BOLD data and are practically equivalent to state-of-the-art correction methods which require reverse phase encoded data. Future directions include additional validation studies, integration with other preprocessing operations, retraining with broader pathologies, and investigating the effects of spin echo versus gradient echo images for training and distortion correction. In summary, we demonstrate SynBOLD-DisCo corrects distortion of fMRI when reverse phase encoding scans or field maps are not available.
Introduction
Functional Magnetic Resonance Imaging (fMRI) is a cornerstone of noninvasive brain research, relying on the Blood Oxygen Level Dependent (BOLD) signal. However, the Echo Planar Imaging (EPI) sequences used for fMRI are highly sensitive to magnetic field inhomogeneities. These inhomogeneities cause geometric distortions, particularly near air-tissue interfaces like the sinuses, leading to significant mismatches between functional data and anatomical structural images.
Standard correction methods, such as FSL’s topup, require additional calibration data—either field maps or reverse phase encoded images. Unfortunately, many legacy or clinical datasets lack these scans, rendering state-of-the-art distortion correction impossible. This paper introduces SynBOLD-DisCo, a deep-learning-based pipeline designed to enable high-quality distortion correction for such limited datasets.
Methodology
SynBOLD-DisCo leverages a 3D U-Net to synthesize undistorted BOLD images from standard T1-weighted (T1w) structural scans. The pipeline operates in three distinct stages:
- Preprocessing: Distorted fMRI data and T1w structural images are standardized, normalized, and registered to a common atlas space.
- Synthesis: A 3D U-Net takes the distorted BOLD and undistorted T1w images as inputs to generate a "synthetic undistorted BOLD" target. The network is trained using a Siamese-like loss function to ensure consistency across different distortion directions.
- Distortion Correction: Instead of using the synthetic image directly, the pipeline uses it as a geometric target for FSL’s
topup. By treating the synthetic image as a zero-distortion reference,topupcalculates a field map that corrects the entire fMRI time series.
Evaluation and Results
The model was trained and validated on a diverse dataset comprising 265 subjects across 9 different sites, ensuring robustness across various scanner manufacturers and resolutions. Quantitative evaluation demonstrated that SynBOLD-DisCo is practically equivalent to the gold standard (using real reverse phase encoded data) in terms of mean absolute percent difference.
Key findings include:
- Geometric Accuracy: Corrected images showed significantly higher mutual information with structural T1w images compared to raw distorted data.
- Robustness: The use of an ensemble of five models (from 5-fold cross-validation) improved synthesis reliability and reduced hallucinated features.
- Generalizability: External validation on a separate NIH cohort confirmed the pipeline's effectiveness on real-world "limited" datasets acquired without reverse phase encoding.
Discussion and Conclusion
While SynBOLD-DisCo provides a powerful solution for missing calibration scans, some limitations remain. The model was primarily trained on Anterior-to-Posterior (AP) and Posterior-to-Anterior (PA) phase encoding directions; its performance on other directions requires further study. Additionally, while the tool corrects geometric warping, it cannot recover signal lost to "dropout" in areas of extreme inhomogeneity. Future work will focus on integrating the pipeline into standard preprocessing suites like fMRIPrep and expanding training to include diverse brain pathologies.
In summary, SynBOLD-DisCo successfully bridges the gap between deep learning and physics-based correction tools. By synthesizing the necessary calibration targets, it unlocks state-of-the-art distortion correction for historical and clinical fMRI data that would otherwise remain uncorrected.
Summary updated