[Paper Review] SUGAR: Spherical Ultrafast Graph Attention Framework for Cortical Surface Registration

Jianxun Ren, Ning An+13 more|arXiv (Cornell University)|Jul 2, 2023

Acute Ischemic Stroke ManagementMedicine3 citations

TL;DR

SUGAR is a novel, unsupervised deep learning framework for cortical surface registration that uses a U-Net-based spherical graph attention network with Euler angle-based deformation modeling. It achieves sub-second processing per subject—12,000× faster than conventional methods—while matching or exceeding state-of-the-art accuracy and distortion control across 10,000+ scans from 7 datasets.

ABSTRACT

Cortical surface registration plays a crucial role in aligning cortical functional and anatomical features across individuals. However, conventional registration algorithms are computationally inefficient. Recently, learning-based registration algorithms have emerged as a promising solution, significantly improving processing efficiency. Nonetheless, there remains a gap in the development of a learning-based method that exceeds the state-of-the-art conventional methods simultaneously in computational efficiency, registration accuracy, and distortion control, despite the theoretically greater representational capabilities of deep learning approaches. To address the challenge, we present SUGAR, a unified unsupervised deep-learning framework for both rigid and non-rigid registration. SUGAR incorporates a U-Net-based spherical graph attention network and leverages the Euler angle representation for deformation. In addition to the similarity loss, we introduce fold and multiple distortion losses, to preserve topology and minimize various types of distortions. Furthermore, we propose a data augmentation strategy specifically tailored for spherical surface registration, enhancing the registration performance. Through extensive evaluation involving over 10,000 scans from 7 diverse datasets, we showed that our framework exhibits comparable or superior registration performance in accuracy, distortion, and test-retest reliability compared to conventional and learning-based methods. Additionally, SUGAR achieves remarkable sub-second processing times, offering a notable speed-up of approximately 12,000 times in registering 9,000 subjects from the UK Biobank dataset in just 32 minutes. This combination of high registration performance and accelerated processing time may greatly benefit large-scale neuroimaging studies.

Motivation & Objective

• Address the critical need for computationally efficient and accurate cortical surface registration in large-scale neuroimaging studies.
• Overcome the limitations of conventional registration methods, which are too slow for large datasets despite high accuracy.
• Bridge the gap between deep learning’s representational power and practical deployment by achieving both high speed and high fidelity in registration.
• Develop a unified framework that supports both rigid and non-rigid registration on spherical cortical surfaces.
• Minimize topological distortions and geometric inconsistencies during registration using specialized loss functions and data augmentation.

Proposed method

• Proposes a U-Net-based spherical graph attention network to model cortical surface geometry using graph convolutional layers on spherical meshes.
• Represents deformation fields using Euler angles to ensure smooth, differentiable, and interpretable non-rigid transformations.
• Introduces a multi-component loss function combining similarity loss, fold loss, and multiple distortion losses to preserve topology and minimize geometric distortion.
• Employs a novel data augmentation strategy tailored for spherical surfaces, enhancing generalization and robustness during training.
• Uses an unsupervised training paradigm with no ground-truth deformation labels, relying solely on image similarity and geometric regularization.
• Applies spherical harmonic-based preprocessing to stabilize the input representation and improve training convergence.

Experimental results

Research questions

• RQ1 Can a deep learning-based framework achieve both sub-second processing time and state-of-the-art registration accuracy on cortical surface data?
• RQ2 How effective is the proposed spherical graph attention network in capturing complex cortical morphology compared to conventional methods?
• RQ3 To what extent do the fold and distortion losses improve topological fidelity and geometric accuracy in non-rigid registration?
• RQ4 Can the proposed data augmentation strategy significantly enhance model generalization across diverse neuroimaging datasets?
• RQ5 How does SUGAR perform in test-retest reliability and consistency across multiple large-scale datasets?

Key findings

• SUGAR achieves a processing time of under one second per subject, representing a ~12,000× speed-up over conventional methods when registering 9,000 UK Biobank subjects in 32 minutes.
• The framework demonstrates comparable or superior registration accuracy to state-of-the-art conventional and learning-based methods across 7 diverse datasets.
• The inclusion of fold and distortion losses significantly reduces topological errors and geometric distortions, improving registration reliability.
• Test-retest reliability analysis shows high consistency across repeated scans, confirming robustness and stability of the learned transformations.
• The proposed data augmentation strategy improves registration performance by reducing overfitting and enhancing generalization across different imaging protocols and subjects.
• SUGAR maintains high performance across diverse populations and acquisition protocols, indicating strong generalization capability in real-world neuroimaging applications.