Segmentation of radiographs of hands with joint damage using customized active appearance models

This research is part of a project that investigates the possibilities of automating the assessment of joint damage in hand radiographs. The goal of this research is to design a robust segmentation algorithm for the hand skeleton. The algorithm is based on active appearance models (AAM) [1], which have been used for hand segmentation before by [2]. The results will be used in future research for radiographic assessment of rheumatoid arthritis and the early detection of joint damage. New in this work with respect to [2] is the use of multiple object warps for each individual bone in a single AAM. This method prevents modeling and reconstruction defects caused when warping overlapping objects. This makes the algorithm more robust in cases where joint damage is present. The current implementation of the model includes the metacarpals, proximal phalanges, and the carpal region. For a first experimental evaluation a collection of 50 hand radiographs has been gathered. The image collection was split into a training set (40) and a test set (10) in order to test the algorithm’s performance. First results show that in 8 images from the test set the bone contours are detected correctly within 1.3 mm (1 STD) at 15 pixels/cm resolution. In two images not all contours are detected correctly. Possibly this is caused by extreme deviations in these images that have not yet been incorporated in the model due to a limited training set. A larger dataset is necessary for optimizing the AAM and improving the quality and reliability of the results. [1] Cootes, T.F., Edwards, G.J., and Taylor, C.J., "Active appearance models," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 23, no. 6, pp. 681-685, 2001. [2] Thodberg, H.H. Hands-on Experience with Active Appearance Models. Sonka and Fitzpatrick. Medical Imaging 2002: Image Proceedings 4684, 495-506. 2002. SPIE.