GPU-Accelerated RSF Level Set Evolution for Large-Scale Microvascular Segmentation

Abstract

Microvascular networks are challenging to reconstruct because they are composed of individual structures near the resolution limit of existing microscopes, making boundary detection difficult. Machine learning tools like convolutional neural networks are effective at semantic segmentation, however they cannot leverage existing information about the global structure of these networks. Level sets are suitable for solving this problem, since they can integrate both the local tube-like structure of capillaries and globally complex topology. However active contours are computationally intensive, making them impractical for terabyte-scale images. We propose a highly parallel formulation of the region-scalable fitting (RSF) model, making it viable for three-dimensional gigavoxel images. We first train a U-Net to provide an initial contour, and then use the proposed RSF model to finalize the segmentation. We tested this approach on microvascular data acquired using multiple state-of-the-art imaging methods. We assess the performance using a Monte-Carlo validation technique and compare results to existing algorithms. This study showcases the practical application of the RSF model, emphasizing its utility in the challenging domain of large-scale high-topology network segmentation with a particular focus on building microvascular models.

Resources