Recon-all-clinical: Cortical surface reconstruction and analysis of heterogeneous clinical brain MRI

TL;DR

Recon-all-clinical tackles the challenge of cortical surface analysis on heterogeneous clinical brain MRI by marrying a deep CNN that predicts signed distance functions with traditional geometry processing to place topologically correct WM and pial surfaces. The method uses domain randomization to train on synthetic MRI data derived from high-quality segmentations, enabling zero retraining for new acquisitions and contrasts. It demonstrates robust cortical reconstruction, parcellation, and thickness estimation across diverse modalities and resolutions, validated on large clinical datasets and test-retest data, and outputs FreeSurfer-compatible results for seamless integration into existing workflows. This approach broadens the accessibility of detailed cortical analyses to vast clinical archives, with implications for studying aging, neurodegenerative diseases, and underrepresented populations where high-quality research scans are scarce.

Abstract

Surface-based analysis of the cerebral cortex is ubiquitous in human neuroimaging with MRI. It is crucial for cortical registration, parcellation, and thickness estimation. Traditionally, these analyses require high-resolution, isotropic scans with good gray-white matter contrast, typically a 1mm T1-weighted scan. This excludes most clinical MRI scans, which are often anisotropic and lack the necessary T1 contrast. To enable large-scale neuroimaging studies using vast clinical data, we introduce recon-all-clinical, a novel method for cortical reconstruction, registration, parcellation, and thickness estimation in brain MRI scans of any resolution and contrast. Our approach employs a hybrid analysis method that combines a convolutional neural network (CNN) trained with domain randomization to predict signed distance functions (SDFs) and classical geometry processing for accurate surface placement while maintaining topological and geometric constraints. The method does not require retraining for different acquisitions, thus simplifying the analysis of heterogeneous clinical datasets. We tested recon-all-clinical on multiple datasets, including over 19,000 clinical scans. The method consistently produced precise cortical reconstructions and high parcellation accuracy across varied MRI contrasts and resolutions. Cortical thickness estimates are precise enough to capture aging effects independently of MRI contrast, although accuracy varies with slice thickness. Our method is publicly available at https://surfer.nmr.mgh.harvard.edu/fswiki/recon-all-clinical, enabling researchers to perform detailed cortical analysis on the huge amounts of already existing clinical MRI scans. This advancement may be particularly valuable for studying rare diseases and underrepresented populations where research-grade MRI data is scarce.

Figures (7)

• Figure 1: Outputs of our recon-all-clinical pipeline: The pipeline accepts an MRI scan of any contrast and resolution as input and generates multiple outputs similar to the recon-all Freesurfer pipeline fischl1999cortical: (a) Volumetric Labeling and T1 contrast synthesis and super-resolution obtained from SynthSegbillot2023synthseg and SynthSR iglesias2023synthsr, (b) Distance prediction and Surface Extraction, (c) Surface Atlas Registration, and (d) Parcellation, followed by the computation of statistical analysis in a common coordinate frame.
• Figure 2: Overview of our proposed approach for cortical analysis of clinical brain MRI scans of any resolution and MRI contrast, without retraining. Training data (a) consisting of isotropic signed distance maps (b) and label maps used to generate the infinite random synthetic data (c) with diverse resolutions, contrasts, and orientations via a generative model. The 3D U-Net is then trained on these data to predict isotropic distance maps, employing L2 loss for error measurement between the training signed distance maps (b) and the predicted maps (d). For testing, for any MRI scans are input (e), the trained 3D U-Net generates predictions of the isotropic SDFs (f) for scans of any resolution and contrast. (g) Subsequent geometry processing of these predictions yields topologically correct cortical surfaces, as well as parcellation and thickness measures.
• Figure 3: Impact of various loss functions across T1, T2, and FLAIR modalities at multiple resolutions. The boxplots represent the distribution of thickness difference at resolutions ranging from 1 mm to 6 mm. Each modality is analyzed with respect to different loss functions, demonstrating the variability and sensitivity of reconstruction accuracy to the choice of loss function within each MRI modality.
• Figure 4: Error distribution in cortical thickness estimation across parcels: Boxplots represent the deviation in estimated cortical thickness measurements obtained by the recon-all and the recon-all-clinical method, using the MIRIAD test-retest dataset. The y-axis indicates the error magnitude in millimeters, with separate color coding for each method. The analysis compares results over two different timepoints across various cortical parcels named on the x-axis. recon-all-clinical method demonstrates a lower mean error and, hence, higher reliability compared to the recon-all, particularly relevant for longitudinal studies due to its consistent accuracy over time. The median error is denoted by the central line within each box, while the edges of the box represent the interquartile range. The whiskers extend to the most extreme data points not considered outliers, which are plotted individually.
• Figure 5: Mean cortical thickness across parcels as a function of age: The first row depicts the relationship between age and mean cortical thickness derived from recon-all processing of 1 mm MPRAGE T1 volumes. The second row presents data from recon-all-clinical, applied to clinical acquisitions with variable resolution, direction, and modality. Each plot corresponds to one of six brain parcels (representative of aging): Precentral, ParsTriangularis, Supramarginal, InferiorParietal, Pericalcarine, and Cuneus. The y-axis represents mean cortical thickness, while the x-axis denotes age. The trend lines are built independently for each region with a B-spline model with three control points, with linear correction for gender and scan resolution.