Fully automatic segmentation of wrist bones for arthritis patients

A reliable method to evaluate and follow the course of arthritis is given by examination of the carpal bones within the wrist joint. Humans typically have eight such small angular bones arranged in two rows. The small size as well as the number make manual segmentation for an analysis of the disease progression a tedious process. Further, fully automatic approaches are still not very reliable. To support medical treatment we present a fully automatic machine learning approach which (i) finds a bounding box around every bone and (ii) outlines the contour and computes a 3-D model of every carpal. The proposed approach has been successfully evaluated on 110 clinical wrist data sets of arthritis patients. The data consists of 59 T1 and 51 T2 weighted MRI images. With the point-to-mesh error deviating from ground truth an average of 0.48 ± 0.45 mm / 0.59 ± 0.49 mm on T1 / T2 modality, accurate segmentation results have been achieved.