Abstract
To achieve fast, robust, and accurate reconstruction of the human cortical surfaces from 3D magnetic resonance images (MRIs), we develop a novel deep learning-based framework, referred to as SurfNN, to reconstruct simultaneously both inner (between white matter and gray matter) and outer (pial) surfaces from MRIs. Different from existing deep learning-based cortical surface reconstruction methods that either reconstruct the cortical surfaces separately or neglect the interdependence between the inner and outer surfaces, SurfNN reconstructs both the inner and outer cortical surfaces jointly by training a single network to predict a midthickness surface that lies at the center of the inner and outer cortical surfaces. The input of SurfNN consists of a 3D MRI and an initialization of the midthickness surface that is represented both implicitly as a 3D distance map and explicitly as a triangular mesh with spherical topology, and its output includes both the inner and outer cortical surfaces, as well as the midthickness surface. The method has been evaluated on a large-scale MRI dataset and demonstrated competitive cortical surface reconstruction performance.
Introduction
Mapping the complex folds of the human cerebral cortex—a process known as Cortical Surface Reconstruction (CSR)—is a fundamental task in neuroimaging. Accurate surface models are essential for quantifying cortical thickness, surface area, and folding patterns, which serve as critical biomarkers for brain development and neurodegenerative diseases such as Alzheimer’s.
Traditionally, CSR is performed using established software suites like FreeSurfer. While highly accurate, these traditional pipelines are computationally intensive, often requiring several hours per subject and occasional manual intervention. Recent deep learning (DL) approaches have significantly accelerated this process, yet they often struggle with two primary issues: independence bias and architectural complexity. Many DL models reconstruct the inner (white matter) and outer (pial) surfaces independently, ignoring their inherent anatomical coupling. Furthermore, hybrid architectures that combine Convolutional Neural Networks (CNNs) for image processing with Graph Neural Networks (GNNs) for mesh manipulation can be difficult to optimize.
SurfNN addresses these challenges by introducing a joint reconstruction framework that leverages the midthickness surface as an anatomical anchor, ensuring consistency and efficiency.
Methodology
The Midthickness Anchor
SurfNN simplifies the reconstruction problem by focusing on the midthickness surface—the layer residing exactly halfway between the inner and outer boundaries. By centering the reconstruction on this median layer, the model explicitly encodes the interdependence of the cortical boundaries, reducing the search space and ensuring that the inner and outer surfaces do not intersect.
Network Architecture
The framework utilizes a unified Fully Convolutional Network (FCN) with a "Y-shape" architecture. The input to the network is a 3D MRI volume paired with an initial "template" midthickness surface. This template is provided in two formats: an explicit triangular mesh and an implicit Signed Distance Function (SDF) map. The SDF map provides the network with spatial context regarding the distance of each voxel from the initial surface, which significantly aids feature extraction.
The network branches into two specialized paths:
- Deformation Branch: This branch predicts a Stationary Velocity Field (SVF) used to warp the initial template into the precise geometry of the midthickness surface.
- Thickness Branch: This branch predicts the Half Cortical Thickness (HCT) for every point on the reconstructed midthickness surface.
Joint Reconstruction
The final inner and outer surfaces are derived from the predicted midthickness surface and the HCT. By shifting the midthickness vertices inward and outward along their normal vectors by the HCT value, the model generates the white matter and pial surfaces simultaneously. This geometric constraint ensures anatomical plausibility and maintains a realistic relationship between the two boundaries.
Experimental Results
Dataset and Training
SurfNN was evaluated using the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, comprising 747 subjects for training, 70 for validation, and 113 for testing. Due to the lack of a manual 3D gold standard, surfaces generated by FreeSurfer v7.2.0 were utilized as the surrogate ground truth for training and benchmarking.
Performance and Efficiency
SurfNN demonstrated superior performance compared to both explicit methods (e.g., PialNN) and implicit methods (e.g., DeepCSR):
- Accuracy: The model achieved a 55.2% improvement in Average Distance (AD) for pial surfaces compared to PialNN.
- Speed: SurfNN is remarkably efficient, reconstructing both surfaces in approximately 1.8 seconds once the initial segmentation is complete. This is significantly faster than the 3.2 seconds required by cortexODE and the several minutes required by implicit DL methods.
- Robustness: The model maintained high accuracy across varying mesh complexities, tested up to approximately 130,000 vertices.
Discussion and Conclusion
A key insight from the development of SurfNN was the impact of the Signed Distance Function (SDF). Ablation studies revealed that including the SDF improved reconstruction accuracy by nearly 20%, highlighting the importance of providing the network with implicit geometric cues. Furthermore, by remaining within the FCN domain until the final deformation step, SurfNN avoids the computational overhead associated with hybrid CNN-GNN models.
While SurfNN represents a significant advancement, it remains dependent on FreeSurfer-generated ground truth, potentially inheriting systematic biases. Future work will focus on refining the initialization process and applying these high-fidelity thickness maps to clinical studies to improve the detection of early-stage brain atrophy. In conclusion, SurfNN provides a fast, robust, and anatomically consistent framework for large-scale cortical surface reconstruction.
Summary updated