Automatic Landmark Detection and Registration of Brain Cortical Surfaces via Quasi-Conformal Geometry and Convolutional Neural Networks

Yuchen Guo, Qiguang Chen, Gary P. T. Choi, Lok Ming Lui

Introduction

Comparing the highly folded surfaces of human brains—a process known as surface registration—is a fundamental task in medical imaging. It enables researchers to identify structural variations associated with neurological conditions such as Alzheimer’s disease. Traditionally, accurate registration requires experts to manually delineate anatomical landmarks, such as the sulci (valleys) of the brain, followed by computationally intensive optimization algorithms to align these surfaces.

Guo et al. propose an automated pipeline that integrates the mathematical rigor of quasi-conformal (QC) geometry with the efficiency of Convolutional Neural Networks (CNNs). This framework automates landmark detection and surface registration, achieving a processing speed over 90% faster than conventional methods while maintaining high accuracy.

Overview of the automated brain registration pipeline, from input brain to registered 3D model.

Methodology: A Three-Stage Deep Learning Framework

The proposed framework operates on a 2D "disk" representation of the brain surface and consists of three specialized neural networks:

  1. LD-Net (Landmark Detection): This network automatically traces landmark curves based on the brain's mean curvature and user-provided start and end points. By treating curvature as a 2D image, the CNN identifies sulcal valleys much like a human expert.
  2. CP-Net (Coefficient Prediction): This stage predicts the Beltrami coefficient ($\mu$), a complex-valued function in QC geometry that describes mapping distortion. A tanh activation function ensures $|\mu| < 1$, which mathematically guarantees a bijective (fold-free) mapping.
  3. DBS-Net (Disk Beltrami Solver): An unsupervised network that converts the predicted $\mu$ into a 2D mapping using Fourier approximation. This approach focuses on low-frequency components to enhance speed and reduce noise.
The three-stage architecture: Landmark Detection, Coefficients Prediction, and Disk Beltrami Solver.

Data and Training

To overcome the scarcity of labeled medical data, the authors used the ImageNet dataset to generate synthetic Beltrami coefficients for training. Evaluation was performed on 24 human brain cortical surfaces from the OASIS dataset, processed via FreeSurfer.

Key Insights

  • Geometric Constraints: By predicting Beltrami coefficients rather than raw coordinates, the network is physically constrained to produce bijective mappings, preventing the "folding" errors common in unconstrained deep learning models.
  • Fourier Compression: Leveraging the fact that essential brain deformations occur in the low-frequency spectrum, the use of Fast Fourier Transform (FFT) layers significantly improves computational efficiency.

Results and Performance

The framework demonstrates a dramatic improvement in registration speed. While traditional optimized harmonic mapping requires between 7 and 46 seconds per brain, the proposed method completes the task in approximately 0.5 seconds—a reduction of over 90%. Despite this speed, landmark alignment accuracy remains comparable to or better than traditional optimization-based approaches, with minimal geometric distortion.

Conclusion and Future Work

This research presents a significant step toward real-time, automated brain surface analysis. While currently limited to simply-connected open surfaces (disk-like topologies) and requiring manual endpoint selection for sulci, future work will focus on extending the framework to spherical topologies and fully automating endpoint detection. This approach highlights the power of combining classical geometric theory with modern deep learning to solve complex medical imaging challenges.

Summary updated

BibTeX
@article{guo2023automatic,
  author = {Yuchen Guo and Qiguang Chen and Gary P. T. Choi and Lok Ming Lui},
  title = {{Automatic Landmark Detection and Registration of Brain Cortical Surfaces via Quasi-Conformal Geometry and Convolutional Neural Networks}},
  journal = {Computers in Biology and Medicine},
  year = {2023},
  doi = {10.1016/j.compbiomed.2023.107185},
}
Figure