Weakly-Supervised Cortical Surfaces Reconstruction from Brain Ribbon Segmentations

Hao Zheng, Xiaoyang Chen, Hongming Li, Tingting Chen, Peixian Liang, Yong Fan

Keywords: Cortical surface reconstruction, diffeomorphic deformation, ODE, Brain MRI

TL;DR: We propose a weakly supervised method to reconstruct multiple cortical surfaces from brain MRI using ribbon segmentation maps.

Abstract: Deep learning-based cortical surface reconstruction (CSR) approaches typically rely on supervision information provided by pseudo ground truth generated by conventional CSR methods, subject to errors associated with the supervision information and also increasing computational cost of training data preparation. We propose a new method to jointly reconstruct multiple cortical surfaces using weak supervision from brain MRI ribbon segmentation results. Our approach initializes a midthickness surface, which is then deformed inward and outward to form the inner (white matter) and outer (pial) cortical surfaces, respectively, by jointly learning diffeomorphic flows by minimizing loss functions to optimize the surfaces towards the boundaries of the cortical ribbon segmentation maps. Specifically, a boundary surface loss drives the initialization surface to the inner and outer boundaries, while an inter-surface normal consistency loss regularizes the pial surface in challenging deep cortical sulci regions. Additional regularization terms are utilized to enforce edge length uniformity and smoothness of the reconstructed surfaces. Our method has been evaluated on two large-scale adult brain MRI datasets and one infant brain MRI dataset, demonstrating comparable or superior performance in CSR in terms of accuracy and surface regularity compared to alternative supervised deep learning methods.

Primary Area: Machine learning for healthcare

Flagged For Ethics Review: true

Submission Number: 6996