Fast and memory efficient feature detection using multiresolution probabilistic boosting trees

This paper presents a highly optimized algorithm for fast feature detection in 3D volumes. Rapid detection of structures and landmarks in medical 3D image data is a key component for many medical applications. To obtain a fast and memory efficient classifier, we introduce probabilistic boosting trees (PBT) with partial cascading and classifier sorting. The extended PBT is integrated into a multiresolution scheme, in order to improve performance and works on block cache data structure which optimizes the memory footprint. We tested our framework on real world clinical datasets and showed that classical PBT can be significantly speeded up even in an environment with limited memory resources using the proposed optimizations.