Self-Supervised Cortical Surface Reconstruction for Ultra High-resolution ex vivo 7T MRI

Haoxiang Li, Mingxuan Liu, Hongjia Yang, Yi Liao, Haibo Qu, Jonathan R. Polimeni, Qiyuan Tian

Keywords: Ex vivo Brain, 7T MRI, Self-supervised Learning, Cortical Surface Reconstruction

TL;DR: SelfCSR is the first learning-based ex vivo CSR method and offering a novel tool for ex vivo MRI-based neuroimaging research.

Abstract: Ex vivo brain MRI enables sub-millimeter ultra-high-resolution studies, uncovering structural details unattainable with in vivo MRI. Cortical surface reconstruction (CSR) based on these detailed images is crucial for studying cortical anatomy and structure. Despite this potential, methodological development in ex vivo MRI has been constrained by several factors: scarcity of datasets, limited imaging resources, pronounced susceptibility artifacts, and signal inhomogeneity. While learning-based CSR methods have been proposed to accelerate reconstruction processes, they face a fundamental limitation—requiring CSR results from classic methods like FreeSurfer as training references, making them mostly only suitable for in vivo adult MRI data and unsuitable for the unique characteristics of ex vivo brain imaging. To address this challenge, we propose SelfCSR, a self-supervised deep learning framework for accurate ex vivo 7T MRI CSR without the need for manually labeled training data.

Submission Number: 106