MedVoxelHD: Improved CUDA-accelerated morphological Hausdorff distances in medical image analysis

Abstract

The selection of optimal segmentation metrics is a critical aspect in the development of successful algorithms. According to new research, average Hausdorff distance (HD) should be the primary metric that evaluates contour alignment. This article introduces a ready-to-use implementation that is at least an order of magnitude faster than the previous best, with the additional advantage of providing unique and detailed information on per-voxel contribution to the overall HD result. Moreover, the algorithm is implemented as a Python Pytorch extension that can be used with most contemporary medical segmentation packages.

Keywords

• CUDA
• Computer tomography
• Medical image segmentation