Cortical Analysis of Heterogeneous Clinical Brain MRI Scans for Large-Scale Neuroimaging Studies

Karthik Gopinath, Douglas N. Greve, Sudeshna Das, Steven E. Arnold, Colin G. Magdamo, Juan Eugenio IglesiasMICCAI 2023 · DOI

TL;DR

This paper introduces a novel method for analyzing brain MRI scans from clinical settings, which are notoriously diverse in resolution and contrast. By combining a deep learning model trained with domain randomization and a classical geometry processing pipeline, the method can accurately reconstruct cortical surfaces, enabling large-scale studies on underrepresented populations and rare diseases.

What Problem Does This Paper Address?

Analyzing brain MRI scans for research purposes typically relies on high-quality, isotropic scans (like 1mm MPRAGEs) with excellent gray-white matter contrast. However, the vast majority of brain MRIs are acquired for clinical reasons, resulting in scans with varying resolutions, slice spacings, orientations, and contrasts. This heterogeneity makes them unsuitable for standard neuroimaging analysis pipelines, such as FreeSurfer, which are sensitive to these variations. This limitation prevents researchers from leveraging the massive amount of clinical data available for large-scale studies, especially for conditions affecting underrepresented groups or rare diseases. The paper aims to bridge this gap by enabling robust cortical analysis on these heterogeneous clinical scans without requiring retraining for each new acquisition protocol.

Methods

The proposed method consists of two main components:

  1. Learning of Signed Distance Functions (SDFs): A convolutional neural network (CNN) is trained to predict voxel-wise SDFs for the white matter (WM) and pial surfaces. Crucially, this CNN is trained using a domain randomization approach. Synthetic MRI scans are generated on-the-fly with randomized orientations, resolutions, slice spacings, and contrasts, derived from existing 3D segmentations. This technique allows the CNN to learn representations that are invariant to these acquisition parameters, making it robust to unseen clinical data. The CNN predicts SDFs, which implicitly define the surfaces.
  2. Geometry Processing for Surface Placement: After the CNN predicts the SDFs for a clinical scan, a classical geometry processing module is used to reconstruct the WM and pial surfaces. This module leverages existing techniques for mesh processing, including hole filling, tessellation, smoothing, and automated manifold surgery to ensure topological correctness (e.g., preventing self-intersections and maintaining spherical topology). The surfaces are then iteratively deformed by minimizing an objective function that balances fidelity to the predicted SDFs with geometric regularization terms. This ensures that the final surfaces are geometrically accurate and topologically sound.

Datasets

  • Training Data: 150 subjects from the Human Connectome Project (HCP) were used to generate training data. FreeSurfer was run on their isotropic T1 scans to obtain segmentations and surface meshes, which were then used to create synthetic training examples.
  • Evaluation Datasets:
    • ADNI: 1mm MPRAGE and corresponding 5mm axial FLAIR scans from 200 subjects were used to compare the method's performance on research-grade vs. clinical-grade scans.
  • Clinical Dataset: A large, highly heterogeneous dataset of 9,735 scans from 1,367 MRI sessions (5,064 successful reconstructions) was used. This dataset includes a wide range of MRI contrasts and resolutions, representing real-world clinical acquisitions.
  • HCP (for ground truth): 581 1mm MPRAGE scans from the HCP dataset were used as a subset to compare against FreeSurfer results.

Key Insights and Ideas

  • Domain Randomization for Robustness: The core innovation is the use of domain randomization during CNN training. This allows the model to generalize to a wide variety of clinical MRI protocols without needing to be retrained for each specific acquisition. This is a significant departure from previous methods that were often limited to specific scan types.
  • Combining Deep Learning and Classical Geometry: The paper effectively combines the power of deep learning for feature extraction (SDF prediction) with the guarantees of classical geometry processing for ensuring topological and geometric correctness of the reconstructed surfaces. This hybrid approach mitigates the limitations of each individual technique.
  • Clinical Data as a Valuable Resource: By enabling analysis of clinical scans, the method unlocks the potential of millions of existing MRI datasets for large-scale neuroimaging studies, particularly for rare diseases and underrepresented populations.

Key Results

  • Accurate Parcellation: The method achieved high accuracy in cortical parcellation, with mean Dice scores of 0.95 on the ADNI dataset, comparable to FreeSurfer on 1mm MPRAGEs.
  • Discriminating Disease Effects: On the ADNI dataset, the method was able to detect the effects of Alzheimer's disease on cortical thickness, recovering a significant portion of the effect size lost by a competing method (SynthSR). Specifically, it showed better performance in detecting atrophy in temporal regions.
  • Aging Effects: The method successfully detected the expected pattern of cortical thinning with aging, outperforming SynthSR in identifying effects in superior temporal regions.
  • Clinical Data Performance: On the heterogeneous clinical dataset, the method maintained high parcellation accuracy (mean Dice of 0.95) and showed promising results in detecting aging-related thinning, although with increased noise compared to 1mm MPRAGEs.
  • Processing Time: The method takes 2-3 hours per scan, which is dominated by the geometry processing step.

Limitations and Future Work

  • Increased Variability in Clinical Data: While the method successfully reconstructs surfaces from heterogeneous clinical scans, the paper notes increased variability in cortical thickness measurements compared to high-quality MPRAGE scans.
  • Future Work: The authors suggest future work could focus on:
  • Improving the reliability of thickness measurements in heterogeneous scenarios.
  • Modeling geometric covariates (e.g., distance to nearest slice, angle between surface and orientation) to further reduce variability.
  • Speeding up the processing time, potentially by replacing some modules (like spherical registration) with faster learning-based methods.
  • Further investigating false positive findings observed in frontal areas.

Summary updated

BibTeX
@inproceedings{gopinath2023cortical,
  author = {Karthik Gopinath and Douglas N. Greve and Sudeshna Das and Steven E. Arnold and Colin G. Magdamo and Juan Eugenio Iglesias},
  title = {{Cortical Analysis of Heterogeneous Clinical Brain MRI Scans for Large-Scale Neuroimaging Studies}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  doi = {10.1007/978-3-031-43993-3\_4},
}
Figure