From Low Field to High Value: Robust Cortical Mapping from Low-Field MRI

TL;DR

This work tackles the limited accessibility of cortical surface morphometry due to reliance on high-field MRI by introducing recon-any, a domain-randomized, out-of-the-box 3D U-Net that predicts cortical signed distance functions from portable low-field MRI scans. The method generalizes across contrasts, resolutions, and even postmortem tissue, delivering surfaces with strong correspondence to high-field references (e.g., surface-area $r=0.96$, Dice $=0.98$, gray matter volume $r=0.93$, thickness up to $r=0.70$). It achieves robust parcellation and morphometry, with Gray matter volume and surface area remaining highly concordant ($r>0.90$, $<10 ext{%}$ error) across most lobes, while thickness remains the most sensitive metric to acquisition parameters. The approach is validated on in vivo and postmortem LF-MRI, and is integrated into FreeSurfer with public code, offering a practical path toward democratizing cortical surface analysis in resource-limited settings and supporting ex vivo neuroanatomical studies.

Abstract

Three-dimensional reconstruction of cortical surfaces from MRI for morphometric analysis is fundamental for understanding brain structure. While high-field MRI (HF-MRI) is standard in research and clinical settings, its limited availability hinders widespread use. Low-field MRI (LF-MRI), particularly portable systems, offers a cost-effective and accessible alternative. However, existing cortical surface analysis tools are optimized for high-resolution HF-MRI and struggle with the lower signal-to-noise ratio and resolution of LF-MRI. In this work, we present a machine learning method for 3D reconstruction and analysis of portable LF-MRI across a range of contrasts and resolutions. Our method works "out of the box" without retraining. It uses a 3D U-Net trained on synthetic LF-MRI to predict signed distance functions of cortical surfaces, followed by geometric processing to ensure topological accuracy. We evaluate our method using paired HF/LF-MRI scans of the same subjects, showing that LF-MRI surface reconstruction accuracy depends on acquisition parameters, including contrast type (T1 vs T2), orientation (axial vs isotropic), and resolution. A 3mm isotropic T2-weighted scan acquired in under 4 minutes, yields strong agreement with HF-derived surfaces: surface area correlates at r=0.96, cortical parcellations reach Dice=0.98, and gray matter volume achieves r=0.93. Cortical thickness remains more challenging with correlations up to r=0.70, reflecting the difficulty of sub-mm precision with 3mm voxels. We further validate our method on challenging postmortem LF-MRI, demonstrating its robustness. Our method represents a step toward enabling cortical surface analysis on portable LF-MRI. Code is available at https://surfer.nmr.mgh.harvard.edu/fswiki/ReconAny

Figures (16)

• Figure 1: Comparison of a classical cortical surface reconstruction (FreeSurfer's recon-all, top) and our deep learning-based approach (Recon-Any, bottom). Recon-all is optimized for T1 HF-MRI and follows a multi-step volumetric segmentation and surface extraction process. Recon-Any directly predicts cortical surfaces from LF-MRI via signed distance functions (SDFs), enabling accurate surface reconstruction across different MRI contrasts and resolutions. In addition, recon-any also predicts volumetric segmentation and generates T1 contrast at 1 mm resolution.
• Figure 2: Cortical surface reconstruction from LF- and HF-MRI for a sample subject (a) HF T1w MRI (1 mm isotropic) processed with FreeSurfer recon-all serves as the reference surfaces used for all comparisons. (b) LF-MRI T1-weighted scans of varying resolutions (1.6 mm $\times$ 1.6 mm $\times$ 5 mm axial, 2 mm, 3 mm, and 4 mm) processed using recon-any show strong alignment with HF-derived surfaces, with error maps highlighting minor discrepancies primarily in deep sulci. (c) T2-weighted LF-MRI scans demonstrate similarly robust reconstruction performance, particularly at 2 mm and 3 mm isotropic resolutions. (d) Parcellation regions are defined by the Desikan-Killiany atlas desikan2006automated, grouped into anatomical lobes (frontal, parietal, occipital, temporal, cingulate, and insular) for morphometric analysis.
• Figure 3: Cortical surface reconstruction errors across different LF-MRI resolutions and contrasts. Left: Absolute Average Distance (AAD). Right: 90th percentile Hausdorff Distance (HD90). Results are shown for both white matter (WM, blue) and pial (orange) surfaces across low-field T1 and T2 MRI scans with 1.6 mm $\times$ 1.6 mm $\times$ 5 mm axial, 2 mm, 3 mm, and 4 mm isotropic voxel sizes. Each box indicates the interquartile range, with the median shown as a central horizontal line. The results demonstrate that T2 scans consistently show lower errors than their T1 scans. For reference, the best-case surface reconstruction accuracy on 1 mm isotropic HF-MRI (used during training) yielded mean AAD values of 0.365 mm (white) and 0.395 mm (pial), and HD90 values of 0.709 mm and 0.852 mm, respectively.
• Figure 4: Cortical parcellation accuracy. Dice similarity coefficients (DSC) between cortical parcellations derived from LF-MRI and the reference parcellations from 1 mm isotropic T1 scans processed with FreeSurfer. Results are grouped by anatomical lobe (Cingulate+Insula, Frontal, Parietal, Temporal, Occipital) and shown separately for T1 (blue) and T2 (orange) LF-MRI scans. Each box plot summarizes the distribution of DSCs across 15 subjects: the boxes represent the interquartile range, the horizontal line within each box denotes the median, and individual points beyond this range are shown as outliers. The results highlight consistently high overlap (DSC $>$ 0.85) across most cortical regions, with slightly lower performance in smaller or entorhinal parcel.
• Figure 5: Cortical morphometry accuracy. Absolute percentage error between morphometric estimates from LF-MRI and the HF-MRI gold standard across major cortical lobes. Three metrics are shown: gray matter volume (left), surface area (middle), and cortical thickness (right), with separate box plots for T1 (blue) and T2 (orange) scans acquired at 3 mm isotropic resolution. Each box plot summarizes the distribution across 15 subjects: boxes represent the interquartile range and points beyond that are plotted as outliers. Surface area and volume estimates show strong agreement (typically within 10%), while cortical thickness shows higher variability, particularly in the occipital lobe, reflecting the sensitivity of thickness estimation to resolution and contrast.