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

Karthik Gopinath, Douglas N. Greve, Colin G. Magdamo, Steven E. Arnold, Sudeshna Das, Oula Puonti, Juan Eugenio IglesiasMedical Image Analysis 2025 · DOI

TL;DR

This paper introduces "recon-all-clinical," a novel method that enables detailed cortical surface reconstruction and analysis from any clinical brain MRI scan, regardless of resolution or contrast. This significantly expands the utility of existing clinical MRI data for large-scale neuroimaging studies, especially for rare diseases and underrepresented populations.

What Problem Does This Paper Address?

Analyzing the human brain's cerebral cortex using MRI is crucial for understanding brain function, development, and disease. Traditional methods, like FreeSurfer's "recon-all," require high-resolution, isotropic T1-weighted scans with excellent contrast. This requirement excludes the vast majority of clinical MRI scans, which are often acquired with varying resolutions, orientations, and contrasts due to practical considerations like scan time and patient comfort. This limitation hinders large-scale studies and the analysis of diverse patient populations. "Recon-all-clinical" aims to bridge this gap by making advanced cortical analysis accessible to the wealth of heterogeneous clinical MRI data.

Methods

The core of "recon-all-clinical" is a hybrid approach combining deep learning with classical geometry processing.

  • Deep Learning for Distance Prediction: A 3D U-Net is trained on synthetically generated MRI data. This synthetic data is created on-the-fly with diverse resolutions, contrasts, and spatial deformations, allowing the network to learn to predict Signed Distance Functions (SDFs) for the white matter and pial surfaces, even from non-standard scans. This "domain randomization" approach is key to achieving generalization across varied clinical data without retraining.
  • Classical Geometry Processing: Once the SDFs are predicted, standard FreeSurfer routines are employed for surface extraction, topological correction, and refinement. This ensures the generation of anatomically accurate and topologically sound cortical surfaces, leveraging the robustness of established methods.
  • Integration: The pipeline integrates existing tools like SynthSeg (for volumetric segmentation) and SynthSR (for super-resolution visualization) to provide a comprehensive analysis workflow.

Datasets

  • Training Data: 500 T1-weighted MRI scans from the Human Connectome Project (HCP) and 500 from the Alzheimer's Disease Neuroimaging Initiative (ADNI) were used to generate segmentations and surfaces for creating synthetic training data.
  • Evaluation Datasets:
    • Simulated Clinical Dataset (SCD): 30 high-resolution T1 scans (15 from HCP, 15 from ADNI) were downsampled to simulate lower-resolution clinical acquisitions across T1, T2, and FLAIR modalities.
  • Test-retest MIRIAD Dataset: Data from 46 individuals (23 with Alzheimer's, 23 age-matched controls) at two time points were used to assess reproducibility.
  • Clinical Dataset: A large dataset of 19,006 scans from 1367 MRI sessions with varying contrasts and resolutions was used for evaluating performance on real-world clinical data.

Key Insights and Ideas

  • Domain Randomization is Key: The success of "recon-all-clinical" hinges on its ability to generalize to unseen data by training on a vast, synthetically generated distribution of MRI scans. This bypasses the need for retraining for every new acquisition protocol.
  • Hybrid Approach Balances Innovation and Robustness: Combining deep learning for initial distance prediction with established geometric methods for surface refinement provides both speed and topological guarantees, crucial for clinical applications.
  • Clinical Data is Underutilized: The paper highlights a significant gap in neuroimaging research due to the inability to leverage the vast amount of existing clinical MRI data, which this method aims to address.
  • Aging Trends are Detectable: Even with heterogeneous clinical data, the method is sensitive enough to detect age-related cortical thinning, demonstrating its clinical relevance.

Key Results

  • Accurate Reconstruction: "Recon-all-clinical" produces geometrically precise cortical reconstructions across various MRI contrasts (T1, T2, FLAIR) and resolutions, with errors generally remaining below 0.4mm even at 6mm slice thickness.
  • Robust Parcellation: The method consistently achieves high accuracy in cortical parcellation, with mean Dice scores above 0.90.
  • Test-retest Reliability: The method demonstrates high reliability and consistency, crucial for longitudinal studies.
  • Aging Detection: The method successfully captures age-related cortical thinning trends, comparable to standard methods on high-quality data.
  • Speed: While not explicitly quantified as a speedup over traditional methods, the deep learning component is designed for rapid inference, and the overall pipeline is made publicly available for efficient use.

Limitations and Future Work

  • Precision at Low Resolutions: While robust, the precision of cortical thickness measurements can be noisier and show diminished aging effects at very low resolutions (e.g., 6mm slice thickness) compared to high-resolution scans. This is attributed to the challenge of measuring minute changes with large voxels.
  • Future Directions: The authors suggest refining computational methods to enhance robustness and accuracy, potentially incorporating uncertainty estimation. They also aim to explore faster alternatives to classical modules to improve computational efficiency and extend the method for longitudinal studies across all MRI contrasts.

Summary updated

BibTeX
@article{gopinath2025recon,
  author = {Karthik Gopinath and Douglas N. Greve and Colin G. Magdamo and Steven E. Arnold and Sudeshna Das and Oula Puonti and Juan Eugenio Iglesias},
  title = {{"Recon-all-clinical": Cortical surface reconstruction and analysis of heterogeneous clinical brain MRI}},
  journal = {Medical Image Analysis},
  year = {2025},
  doi = {10.1016/j.media.2025.103608},
}
Figure