Fast boosting trees for classification, pose detection, and boundary detection on a GPU

Neil Birkbeck, Michal Sofka, Shaohua Kevin Zhou

Abstract: Discriminative classifiers are often the computational bottleneck in medical imaging applications such as foreground/background classification, 3D pose detection, and boundary delineation. To overcome this bottleneck, we propose a fast technique based on boosting tree classifiers adapted for GPU computation. Unlike standard tree-based algorithms, our method does not have any recursive calls which makes it GPU-friendly. The algorithm is integrated into an optimized Hierarchical Detection Network (HDN) for 3D pose detection and boundary detection in 3D medical images. On desktop GPUs, we demonstrate an 80× speedup in simple classification of Liver in MRI volumes, and 30× speedup in multi-object localization of fetal head structures in ultrasound images, and 10× speedup on 2.49 mm accurate Liver boundary detection in MRI.