Ultra-Fast 3D GPGPU Region Extractions for Anatomy Segmentation

Abstract

Region extractions are ubiquitous in any anatomy segmentation. Region growing is one such method. Starting from an initial seed point, it grows a region of interest until all valid voxels are checked, thereby resulting in an object segmentation. Although widely used, it is computationally expensive because of its sequential approach. In this paper, we present a parallel and high performance alternate for region growing using GPGPU capability. The idea is to approximate region growing requirements within an algorithm using a parallel connected-component labeling (CCL) solution. To showcase this, we selected a typical lung segmentation problem using region growing. In CPU, sequential approach consists of 3D region growing inside a mask, that is created after applying a threshold. In GPU, parallel alternative is to apply parallel CCL and select the biggest region of interest. We evaluated our approach on 45 clinical chest CT scans in LIDC data from TCIA repository. With respect to CPU, our CUDA based GPU facilitated an average performance improvement of 240× approximately. Speed up is so profound that it can be even applied to 4D lung segmentation at 6 fps.