Automatic segmentation of the thumb carpometacarpal joint

Abstract

INTRODUCTION: The first carpometacarpal (CMC) joint is highly susceptible to osteoarthritis (OA) which affects over 15% of adults over age 30, two-thirds of which are women. We believe that morphology of the CMC joint plays a role in the development of CMC OA. Here we present an automated pipeline for creating parametric meshes of the CMC joint from CT-images, for the purpose of creating statistical models of CMC joint morphology. This method involves the novel combination of random forest regression [1] with a parametric mesh-based statistical shape model (SSM) to automatically create CMC meshes which are correspondent across individuals and are therefore suitable for statistical shape analysis. METHODS: A training set of 50 CMC joints were manually segmented from CT images of the hand with a resolution of 0.4x0.4x0.625mm (age range: 18 yrs to 67 yrs; 24 females and 26 males). A custom piecewise parametric template mesh was fitted to each segmented data cloud, resulting in a set of correspondent meshes of the metacarpal and trapezium which were used to train the SSM. 3-D Haar-like features were sampled from the image about each mesh node and used to train a random forest regressor for each node. During segmentation, the mean mesh was initialized near the in-image CMC joint in the center of the CT image. Random forest regressors then predicted the best-matching image positions of the mesh nodes. The mesh was then fitted to the predicted points using deformations permitted by the SSM resulting in a customized mesh of the CMC joint surface. This segmentation was performed on 15 CT images not a part of the training set. RESULTS: 12 of the 15 data sets were successfully segmented with an average surface-to-surface RMS error of ~1.95 mm. The time taken for the pipeline to segment 15 data sets was approximate 30 minutes. DISCUSSION: This pipeline shows promise for automatically collecting a large population of CMC joint morphology for statistical analysis. Eventually, it may be used in a clinical setting where rapid patient specific mesh generation and analysis would be invaluable.