A Curvature-Guided Coarse-to-Fine Framework for Enhanced Whole Brain Segmentation

Whole brain segmentation, which divides the entire brain volume into anatomically labeled regions of interest (ROIs), is a crucial step in brain image analysis. Traditional methods often rely on intricate pipelines that, while accurate, are time-consuming and require expertise due to their complexity. Alternatively, end-to-end deep learning methods offer rapid whole brain segmentation but often sacrifice accuracy due to neglect of geometric features. In this paper, we propose a novel framework that integrates the key curvature feature, previously utilized by complex surface-based pipelines but overlooked by volume-based methods, into deep neural networks, thereby achieving both high accuracy and efficiency. Specifically, we first train a coarse anatomical segmentation model focusing on high-contrast tissue types, i.e., white matter (WM), gray matter (GM), and subcortical regions. Next, we reconstruct the cortical surfaces using the WM/GM interface and compute curvature features for each vertex on the surfaces. These curvature features are then mapped back to the image space, where they are combined with intensity features to train a finer cortical parcellation model. We also simplify the process of cortical surface reconstruction and curvature computation, thereby enhancing the overall efficiency of the framework. Additionally, our framework is flexible and can incorporate any neural network as its backbone. It can serve as a plug-and-play component to enhance the whole brain segmentation results of any segmentation network. Experimental results on the public Mindboggle-101 dataset demonstrate improved segmentation performance with comparable speed compared to various deep learning methods.