Automatic knee cartilage segmentation from multi-contrast MR images using support vector machine classification with spatial dependencies Kunlei Zhang a, ⁎, Wenmiao Lu b , Pina Marziliano a a School of Electrical and Electronic Engineering, Nanyang Technological University, 639798, Singapore b Beckman Institute, University of Illinois, Urbana-Champaign, IL 61822, USA abstract article info Article history: Received 17 December 2012 Revised 28 May 2013 Accepted 10 June 2013 Keywords: Magnetic resonance imaging (MRI) Multi-contrast Segmentation Knee cartilage Discriminative random field (DRF) Support vector machine (SVM) Accurate segmentation of knee cartilage is required to obtain quantitative cartilage measurements, which is crucial for the assessment of knee pathology caused by musculoskeletal diseases or sudden injuries. This paper presents an automatic knee cartilage segmentation technique which exploits a rich set of image features from multi-contrast magnetic resonance (MR) images and the spatial dependencies between neighbouring voxels. The image features and the spatial dependencies are modelled into a support vector machine (SVM)-based association potential and a discriminative random field (DRF)-based interaction potential. Subsequently, both potentials are incorporated into an inference graphical model such that the knee cartilage segmentation is cast into an optimal labelling problem which can be efficiently solved by loopy belief propagation. The effectiveness of the proposed technique is validated on a database of multi-contrast MR images. The experimental results show that using diverse forms of image and anatomical structure information as the features are helpful in improving the segmentation, and the joint SVM-DRF model is superior to the classification models based solely on DRF or SVM in terms of accuracy when the same features are used. The developed segmentation technique achieves good performance compared with gold standard segmentations and obtained higher average DSC values than the state-of-the-art automatic cartilage segmentation studies. © 2013 Elsevier Inc. All rights reserved. 1. Introduction Musculoskeletal diseases and articular disorders are one of the major health problems in causing work disability. In particular, the human knee joint is commonly affected by acute injury or osteoar- thritis (OA) which is mainly characterized by the articular cartilage pathology. Quantitative cartilage measurements such as the thickness, volume and surface area, are required to properly assess the damage to cartilage. Accurate segmentation of the knee cartilage is the key to obtain these measurements. Because a segmentation system has been the essential task for the cartilage assessment, this study is focused on developing cartilage segmentation methods. A recent review of algorithms for medical image segmentation can be found in [1]. Magnetic resonance (MR) imaging is the most effective imaging modality to detect anatomical changes in knee joints for its excellent tissue contrasts. Fig. 1 illustrates a two-dimensional (2D) sagittal fat suppressed (FS) spoiled gradient recall (SPGR) MR image of the knee joint which is composed of three bones (patella, femur and tibia), three corresponding cartilage compartments (patellar cartilage, femoral cartilage and tibial cartilage), and other tissues including muscle, ligament, tendon, menisci and adipose tissue, etc. MR imaging has commonly been used for knee cartilage segmentation and assessment in the literature [2]. In general, fully manual and semi-automatic cartilage segmentation approaches [3–15] for routine clinical use are laborious and time consuming. Moreover, the intra/inter-observer variability complicates the interpretation of the results. On the other hand, automatic cartilage segmentation from MR images is a challenging task due to the thin variable morphology of cartilages, low contrast between cartilages and other soft tissues, MR artifacts, and intensity inhomogeneity (see Fig. 1). In this work, we address the problem of automatic knee cartilage segmentation using a classification approach, in which we utilize a rich set of image features provided by multiple spectral MR images taken with different sequences (referred to as multi-contrast MR images 1 ) and incorporate the image features with the spatial dependences between neighbouring voxels in an inference graphical model. 1.1. Prior works on automatic segmentation of knee cartilage In recent years, a number of automatic cartilage segmentation works have been reported. Most of them used atlas/template registration, optimal graph, statistical or deformable model-based Magnetic Resonance Imaging 31 (2013) 1731–1743 1 They are also called multi-spectral or multi-modal MR images in some works. ⁎ Corresponding author. E-mail address: kunleizhang@gmail.com (K. Zhang). 0730-725X/$ – see front matter © 2013 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.mri.2013.06.005 Contents lists available at ScienceDirect Magnetic Resonance Imaging journal homepage: www.mrijournal.com