Cervical spine mobility analysis on radiographs: A fully automatic approach
Introduction
A wide majority of adults are, or will be, confronted to back problems in their daily life. For instance, low back pain is a leading cause of work absenteeism. Another illustration concerns the cervical spine trauma which represents the majority of spinal lesions. In this context, whiplash injuries are probably the most recurrent cause for insurance claims but with no well-established diagnosis [1].
Many of these problems are the consequence of an abnormal spinal motion. One practical device to extract spine motion in real-time is the digital videofluoroscopy (DVF). Proposed for the first time in [2], this technique consists in recording X-ray images as video frame images. At that time, the authors showed that it was possible to acquire and study the lumbar spine motion with such a system.
The initial step for studying spine motion in a DVF system is landmarking or locating the vertebræ among the frames. Initially, this stage was operated manually. Nowadays, the contributions in the literature attempt to automate as much as possible the landmarking or vertebra locating procedures. Common approaches are based on template matching. The idea is to build a template defining a vertebra, to place the template on the first frame of the video sequence (generally, this step requires a human intervention) and finally to track the template frame by frame. In this context, approaches based on cross-correlation are presented in the literature. The idea is to define a measure of correlation which can be regarded as a similarity value. In [3], Bifulco et al. used the cross-correlation for automatic recognition of vertebra landmarks. Muggleton and Allen presented a similar approach in [4], but refined it with a second stage where an annular template enclosing the vertebral body edges is used. The cross-correlation has also been considered by Cerciello et al. in [1]. Recently, they proposed an improved version of template matching in [5]. The vertebra edges are enhanced by estimating image gradient. However, they made the observation that fluoroscopic images are very noisy and that gradient estimation is precisely sensitive to noise. Therefore, this issue is solved by a noise suppression technique preserving edges.
Other methods aimed at extracting vertebral boundaries and tracking them in a video sequence are presented. In this context, Zheng et al. integrated Fourier descriptors with Hough transform algorithm in [6]. They obtained promising results on a calibration model but the approach was validated on only 2 real sequences. Another interesting method was developed in [7], where Lam et al. presented a tracking algorithm based on a dynamic Bayesian network with a particle filter at each node. In this framework, prior information of kinematics and anatomical configuration are combined with the information collected from each frame to estimate the present state. Finally, Reinartz et al. argued in [8] that exact contour extraction is not necessary and proposed a method based on the normalized gradient field for tracking vertebræ.
All these contributions in the context of DVF are of great interest. However, conventional X-ray radiography is the most common modality used in emergency rooms since it is relatively inexpensive and fast. Moreover, it remains nowadays the most common method to analyze spinal motion in two dimensions [7]. Surprisingly, the literature on the subject does not focus on automating approaches. Computer-aided protocols such as [9] or [10] require the intervention of an operator on the images. A semi-automatic approach is proposed in [11].
Actually, works on fully automatic vertebra extraction on radiographs are few in the literature. Zamora et al. tried to take advantage of the Generalized Hough Transform (GHT) in [12] but they present a segmentation rate equal to 47% for lumbar vertebræ without providing information about the detection rate. Recently, Dong and Zheng [13] have proposed a method combining GHT and the minimal intervention of a user with only 2 clicks in the image. A fully automatic approach has been developed by Casciaro and Massoptier in [14]. They use a shape constraint characterization by looking for every shape that could be an inter-vertebral disc. Very recently, we presented in [15] an approach to locate the anterior vertebra corners in a radiograph. Nevertheless, these landmarks are not sufficient in order to analyze the cervical spine motion. Therefore, this approach can be combined to a segmentation procedure.
Our contribution is to propose a fully automatic framework for cervical spine mobility analysis on radiographs. Our framework is based on three main steps: fully automatic vertebra detection, vertebra segmentation and vertebra mobility analysis. More precisely, the vertebræ on a given radiograph are detected by the anterior corners. This information is used to initialize a segmentation procedure based on Active Shape Model [16]. Once the exact contour of the vertebræ is extracted, a protocol for cervical spine mobility analysis on radiographs can be performed. To the best of our knowledge, this work is the first report on the subject, since only a few semi-automatic methods have been proposed.
The rest of the paper is organized as follows. Section 2 presents the approach for fully automatic vertebra detection on radiographs. Some aspects of our previous work [15] are adapted to deal with mobility analysis. In Section 3, we describe how the vertebra detection is used to initialize a segmentation procedure based on Active Shape Model. Section 4 reports the detection, segmentation and mobility results for our framework. Section 5 concludes the paper.
Section snippets
Vertebra detection
The vertebra detection is operated with a sequence of 3 stages: interest point detection, interest point description and interest point classification.
Vertebra segmentation
The segmentation methods in the literature are dependent on the imaging modality (e.g. Computed Tomography, X-ray Images, Magnetic Resonance Images, etc.). With regard to X-ray images, the efficient frameworks to segment vertebræ were based on a model. The most common model-based approaches are Active Shape Model (ASM) and Active Appearance Model (AAM), developed by Cootes et al. in, respectively [16], [21]. The former is mainly interested in the shape modelization of a given object while the
Data and methods
The radiographs used in this work were acquired at Jolimont Hospital in Haine-Saint-Paul, Belgium. All the images have been annotated by an experienced radiologist. They were chosen so that all the cervical vertebræ C3–C7 are visible (sometimes partially) with the naked eye. The detection and the segmentation were performed on 245 cervical vertebræ from 49 distinct patients. Thus, the tests are performed on different images and the results are not biased. Patient's age ranges from 35 to 60
Conclusion
The objective of this paper was to develop a framework dedicated to the fully automatic cervical spine mobility analysis on X-ray images. To this aim, we proposed an approach based on three main steps: fully automatic vertebra detection, vertebra segmentation and angular measurement.
The vertebra detection is based on a learning method, where the vertebræ are detected by their anterior corners. General interest points are first detected in the radiograph, then are described with features such as
Conflict of interest
None.
Acknowledgements
We would like to thank Paul Desclée, Physician in the Radiology Department of the Jolimont Hospital, for having provided and annotated the radiographs. The validation of our approach would not have been possible without his help. We are also grateful to Aloys du Bois d’Aische for insightful discussions on this work. Finally, we would like to thank the anonymous reviewers for their valuable comments helping to improve the paper.
Fabian Lecron received the computer science engineering degree from the Faculté Polytechnique de Mons, Belgium, and the management science degree from the Facultés Universitaires Catholiques de Mons (FUCaM), Belgium, respectively in 2008 and 2011. He is now preparing a PhD thesis in the field of statistical modeling of the spine applied to medical image analysis at the University of Mons (formerly Faculté Polytechnique de Mons), Belgium. His main research areas are computer vision, image
References (31)
- et al.
Spine kinematics: a digital videofluoroscopic technique
J Biomed Eng
(1989) - et al.
Automatic location of vertebrae in digitized videofluoroscopic images of the lumbar spine
Med Eng Phys
(1997) - et al.
Advanced template matching method for estimation of intervertebral kinematics of lumbar spine
Med Eng Phys
(2011) - et al.
Automated tracking in digitized videofluoroscopy sequences for spine kinematic analysis
Image Vision Comput
(2009) - et al.
Sagittal plane segmental motion of the cervical spine. A new precision measurement protocol and normal motion data of healthy adults
Clin Biomech
(2002) - et al.
Active shape models-their training and application
Comput Vision Image Underst
(1995) - et al.
Speeded-up robust features (SURF)
Comput Vision Image Underst
(2008) - et al.
On the creation of a segmentation library for digitized cervical and lumbar spine radiographs
Comput Med Imaging Graph
(2011) - et al.
Semiautomatic segmentation of vertebrae in lateral X-rays using a conditional shape model
Acad Radiol
(2007) - et al.
Automatic vertebra tracking through dynamic fluoroscopic sequence by smooth derivative template matching
Automatic recognition of vertebral landmarks in fluoroscopic sequences for analysis of intervertebral kinematics
Med Biol Eng Comput
Automated segmentation of lumbar vertebrae in digital videofluoroscopic images
IEEE Trans Med Imaging
Cervical vertebrae tracking in video-fluoroscopy using the normalized gradient field
Med Image Comput Comput Assist Interv
Segmental vertebral motion in the assessment of neck range of motion in whiplash patients
Int J Legal Med
X-ray image segmentation for vertebral mobility analysis
Int J Comput Assist Radiol Surg
Cited by (8)
Motion analysis of the cervical spine during extension and flexion: Reliability of the vertebral marking procedure
2018, Medical Engineering and PhysicsCitation Excerpt :This new evidence alters the current perception of vertebral motion. Previous range of motion calculation studies used manual and automated analysis programs [12,15–19], which subtracted individual joint motions between the vertebrae [2,5,11,12]. Their validity depends on image quality, degree of motion blurring, precision of vertebral tacking during motion, and natural variability among the subjects.
Investigation of the reproducibility and reliability of sagittal vertebral inclination measurements from MR images of the spine
2014, Computerized Medical Imaging and GraphicsCitation Excerpt :Alternatively, measuring the inclination of vertebral body walls resulted in posterior [8] and anterior [9] tangents. Although the measurements are mostly performed in two-dimensional (2D) sagittal radiographs, which are widely available and adequate due to the visibility of bone structures [10], magnetic resonance (MR) has gained acceptance in spine imaging by providing high-quality three-dimensional (3D) images by a correct selection of imaging parameters [11,12]. As a result, MR images of the spine were already used to measure various vertebral parameters [13–15] and may be helpful in decision making related to the surgical treatment of spinal deformities [16], while the related findings can potentially be used as predictors of clinical prognoses and outcomes [17,18].
Automated Measurement of Lumbar Lordosis on Radiographs Using Machine Learning and Computer Vision
2020, Global Spine JournalAutomatic Detection of Cervical Vertebral Landmarks for Fluoroscopic Joint Motion Analysis
2019, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Fabian Lecron received the computer science engineering degree from the Faculté Polytechnique de Mons, Belgium, and the management science degree from the Facultés Universitaires Catholiques de Mons (FUCaM), Belgium, respectively in 2008 and 2011. He is now preparing a PhD thesis in the field of statistical modeling of the spine applied to medical image analysis at the University of Mons (formerly Faculté Polytechnique de Mons), Belgium. His main research areas are computer vision, image processing and pattern recognition
Mohammed Benjelloun received the electromechanic engineering degree and the management and computer engineering degree respectively in 1986 and 1989. He completed his PhD in applied sciences in 1994 from the Faculty of Engineering, Mons, Belgium (FPMs). He is currently associate professor in the Department of Computer Science at the Faculty of Engineering FPMs of the University of Mons. His research interests include computer vision, image processing, 2D and 3D retrieval, parallel computing and software engineering. He was involved in several projects including industrial and academic researchers and he participates in numerous scientific and academic activities.
Saïd Mahmoudi received in 1997 the B.S. degree in computer science engineering from the University of Science and Technology of Oran (SENIA), Algeria, and in 1999 the M.S. degree in computer science from the LIFL Laboratory, University of Lille 1, France. After that, he joined the FOX-MIIRE Group of the LIFL Laboratory, at the University of Science and Technology of Lille 1, France, where he started his Ph.D. thesis. During this period, he worked on image recognition and 3D retrieval using characteristic views. In 2003, Dr. Mahmoudi obtained the Ph.D. degree in the Computer Science at the University of Science and Technology of Lille 1, France. Between 2003 and 2005 he was associate lecturer at the University of Lille 3, France. Since September 2005, he is associate professor at the Faculty of Engineering of the University of Mons, Belgium. His research interests include medical images processing and computer aided diagnosis, 2D and 3D retrieval and indexing.