Contents lists available at ScienceDirect Pattern Recognition journal homepage: www.elsevier.com/locate/pr Automated segmentation and area estimation of neural foramina with boundary regression model Xiaoxu He a,b, ⁎ , Andrea Lum a,b , Manas Sharma a,b , Gary Brahm a,b , Ashley Mercado a,b , Shuo Li a,b, ⁎ a Digital Imaging Group (DIG), London, ON, Canada N6A 4V2 b Department of Medical Imaging, University of Western Ontario, London, ON, Canada N6A 3K7 ARTICLE INFO Keywords: Automated segmentation Area estimation Neural foramina stenosis Boundary regression model Multiple output support vector regression Multiple kernel learning ABSTRACT Accurate segmentation and area estimation of neural foramina from both CT and MR images are essential to clinical diagnosis of neural foramina stenosis. Existing clinical routine, relying on physician's purely manual segmentation, becomes very tedious, laborious, and inefficient. Automated segmentation is highly desirable but faces big challenges from diverse boundary, local weak/no boundary, and intra/inter-modality intensity inhomogeneity. In this paper, a novel boundary regression segmentation framework is proposed for fully automated and multi-modal segmentation of neural foramina. It creatively formulates the segmentation task as a boundary regression problem which models a highly nonlinear mapping function from substantially diverse neural foramina images directly to desired object boundaries. By leveraging a seamless combination of multiple output support vector regression (MSVR) and multiple kernel learning (MKL), the proposed framework enables the domain knowledge learning in a holistic fashion which successfully handles the extreme diversity posing a tremendous challenge to conventional segmentation methods. The performance evaluation was conducted on a dataset including 912 MR images and 306 CT images collected from 152 subjects. Experimental results show that the proposed automated segmentation framework is highly consistent with physician with average DSI (dice similarity index) as high as 0.9005 (CT), 0.8984 (MR), 0.8935 (MR+CT) and BD (boundary distance) as low as 0.6393 mm (CT), 0.6586 mm (MR), 0.6881 mm (MR+CT). Based on this accurate automated segmentation, the estimated areas, highly correlated to their independent ground truth, have been achieved with correlation coefficient: 0.9154 (CT) and 0.8789 (MR). Hence, the proposed approach enables an efficient, accurate and convenient tool for clinical diagnosis of neural foramina stenosis. 1. Introduction Neural foramina stenosis (NFS), clinically defined as the narrowing of the bony exit (see Fig. 1(a)) of the spinal nerve root, is caused by abnormalities in vertebral and intervertebral disc, such as a decrease in the height of an intervertebral disc, or osteoarthritic changes in the facet joints [1,2]. Symptoms of NFS are very common, affecting up to 80% of the population worldwide, and may cause pain, disability and economic loss [3–5]. For example, each year more than 400,000 Americans suffer from lower back or leg pain [6,7]. Diagnosis and treatment of NFS, often require segmentation of neural foramina images from multiple imaging modalities for estimating its area as quantitative analysis [1,8–10]. Here, MR and CT imaging are often simultaneously required as MR is better to display the stenosis caused by disc abnormality and CT is better to display the stenosis caused by vertebra abnormality (as shown in Fig. 1(b)). For efficient diagnosis and timely treatment of NFS, manual segmentation by physician is bound to be infeasible for neural foramina images because of its known tediousness, inefficiency, and inconsistency [8,10]. Computer processing methods are highly desirable, but face big challenges due to the following complexities in segmentation of neural foramina (as shown in Fig. 1(c)): 1. Complex appearance inhomogeneity: Two types of appearance inhomogeneity are included: (1) Inter-modality intensity difference: In different modalities, the intensity profile of neural foramina is completely different [10]. (2) Intra-modality appearance variation: Even for one specific modality, the structures passing neural foramina are inhomo- geneous and this inhomogeneity varies with different subjects, positions, and spine abnormalities [3]. 2. Great boundary variations: Two types of boundary variations are included: (1) Diverse boundary shape variation: The boundary shape of http://dx.doi.org/10.1016/j.patcog.2016.09.018 Received 29 December 2015; Received in revised form 5 September 2016; Accepted 21 September 2016 ⁎ Corresponding authors. Department of Medical Imaging, University of Western Ontario, London, ON, Canada N6A 3K7. E-mail addresses: xhe244@uwo.ca (X. He), slishuo@gmail.com (S. Li). Pattern Recognition xx (xxxx) xxxx–xxxx 0031-3203/ © 2016 Elsevier Ltd. All rights reserved. Available online xxxx Please cite this article as: He, X., Pattern Recognition (2016), http://dx.doi.org/10.1016/j.patcog.2016.09.018