A fully automatic vertebra segmentation method using 3D deformable fences
Introduction
Spinal surgeries are performed to treat fractures, correct deformity, and replace portions of a spine. Spine fractures can be caused by spine diseases, accidents, or osteoporosis characterized by a reduction in bone density. A deformed spine with an S-shaped curvature-scoliosis-mostly occurs in children and adolescents. Severe cases of osteoarthritis and degenerative disc disease require a spine disc replacement. Such spinal surgeries necessitate segmentation of vertebra in CT images for preoperative planning. However, the complex shape and density variability of a vertebra makes its segmentation challenging.
Segmentation methods can be divided into edge-based, region-based, and integrated methods [1]. Edge-based methods [2], [3] utilize edges to detect a boundary of a region while region-based methods [4] segment a region using its homogeneity. Integrated methods [5] delineate a region boundary using both edge and homogeneity information. Most methods developed for general medical image applications cannot produce accurate results for vertebra segmentation because a vertebra has an ambiguous edge boundary and because it does not contain a homogeneous region. To cope with these characteristics, previous methods intended for vertebra segmentation have incorporated shape information. Statistical shape models (SSMs) [6], [7] and active shape models (ASMs) [8] use shape constraints to overcome ambiguous boundary information. SSMs generate mean shapes using their own shape parameters, such as Fourier and wavelet descriptors, and represent shape variations only with significant descriptors acquired by a principle component analysis (PCA) method. Although SSMs can overcome an ambiguous boundary problem, they cannot converge into an unusual shape or represent small variations in a boundary. The ASM method is a type of SSM that iteratively searches a boundary while maintaining shape constraints and has the same characteristics as an SSM.
Active appearance models (AAMs) [9] incorporate appearance information in addition to shape constraints. With information about texture, AAMs could provide better and more robust results than ASMs in many medical segmentation applications. However, the texture pattern of a vertebral body is so different among patients that its application to vertebra segmentation is difficult. AAMs were applied to vertebra segmentation using only 2D digitized X-ray images, but this application has not been extended to 3D CT images.
A shape constrained deformable model [10] extracts a boundary by minimizing energy including shape energy, which tends to keep a boundary to the model shape. The deformability allows the method to capture an unusual shape and/or shape details. However, it can still provide inaccurate segmentation results because of an incorrect initial condition, such as the position or shape of an initial model. Moreover, shape constraint deformable models consume more computation time for the fitting process than ASMs. A deformable spine model using landmarks [11] attempted to exploit shape information and gray level inhomogeneities using necklace and spring models. The necklace model captures a variation in vertebra structures while the string model represents spinal curvatures. For an initial condition, a user needs to provide a few landmark points to anchor a necklace model. Moreover, the method can be trapped into a local minimum, depending upon the initial condition, and fails to segment abnormal vertebrae, just as most shape based deformable models do.
Interactive tools for spine segmentation [12] were developed to achieve more accurate results. They contain some basic image processing tools and specialized tools for vertebra segmentation. Although the interactive method provides protocols for segmentation, it still requires a laborious manual process.
Much research had also been developed for the segmentation of spinal cord and/or some part of a vertebra. For a curved planar reformation, spine curve parameters were optimized by structural characteristics of vertebra bones [13] but required user interactions to specify start and end vertebrae. An automated spinal column extraction and partition method [14] was developed using routine CT images; in this process, a vertebra region is segmented in a 2D image by fitting a four-part vertebra model. However, the segmentation could not separate the vertebra region into composing vertebra bones, where a spinous process belonging to the upper vertebra exists with a transverse process pertaining to the current vertebra. The authors [15], [16] also developed multiple level set methods to extract only vertebra bodies but did not attempt to segment spinous processes. Moreover, user interactions are required for parameter initializations. Template matching and polar signature were developed to detect boundaries of vertebra bodies [17], [18].
Registration-based methods [19], [20] attempted to segment a vertebra for an intra-operative procedure. Those methods registered a patient vertebra in CT to other image modalities. Thus, they cannot be easily modified for segmentation of vertebra in 3D CT images.
This paper presents a fully automatic vertebra segmentation method using 3D deformable fences for 3D CT images. To the best of our knowledge, it is the first fully automatic method for 3D CT images, and the method may be applied to other image modalities with some modifications. The method extracts the spinal cord and discs along the cord. At each intervertebral disc, a 3D fence separating adjacent vertebrae is generated by a deformable model using 3D valley information. While making a 3D fence, an erroneous curve converged into a local minimum is detected with an evaluation method and corrected by using a minimum cost path finding method. The final vertebra volume is produced by a seed region growing (SRG) method using 3D fences. These methods are described in Section 2. Experimental results of 50 patient data sets are provided in Section 3. Finally, conclusions are drawn in Section 4.
Section snippets
Segmentation method
The proposed method consists of four modules: preprocessing, intervertebral disc search, 3D fence generation and fence-limited labeling. The preprocessing module detects 3D valleys and makes valley-emphasized Gaussian images. Valleys are more suitable features than gradients in separating two adjacent objects because they appear in the middle of two adjacent objects while the gradients appear at the boundaries of each object. The second module automatically extracts a spinal cord, and detects
Results
The proposed method has been applied to 50 patient data sets. The data sets consist of about 170 slices on average and have slice thicknesses between 0.5 mm and 3 mm, with an average thickness of 1.7 mm. Each data set contains 3–8 vertebrae and has an average of 6 vertebrae. Because the correctness of the segmentation depends on the accurate extraction of 3D fences, the correctness of the 3D fence extraction was evaluated by radiologists specializing in vertebra bones. The subjective evaluation
Conclusion
To the best of our knowledge, a fully automatic vertebra segmentation method for 3D CT images has been developed for the first time. The proposed method first extracts the spinal cord and detects intervertebral discs to acquire initial 2D disc passing lines that separate vertebra bodies. From a 2D disc passing line, a 2D curve is then extracted with a deformable model that utilizes 3D valley information and is expanded to form a 3D surface. During expansion, an erroneous 2D curve is detected
Conflict of interest statement
Any people or organizations do not influence our work inappropriately.
Role of the funding source
The Soongsil University has no involvement in the study design, in the collection, analysis and interpretation of data; in the writing of the manuscript; and in the decision to submit the manuscript for publication.
Acknowledgements
This research was supported by the Soongsil University. The spinal data sets used in this study were kindly provided by Prof. Jonghyo Kim of Seoul national university hospital. The authors are grateful to Prof. Dr. Frithjof Kruggel of the University of California, Irvine for comments and proofreading.
Yiebin Kim is a Ph.D. candidate in School of Electronic Engineering at the Soongsil University, Seoul, Korea. He received his BS degree (in 2002) in Department of Electronic engineering and MS degree (in 2004) in Department of Information Telecommunication engineering from Soongsil University. His major research interests include medical image processing, computer vision, and robotics.
References (22)
- et al.
Interactive live-wire boundary extraction
Med Image Anal
(1997) - et al.
Statistical shape model based segmentation of medical images
Comput Med Imaging Graph
(1998) - et al.
Automatic measurement of vertebral shape using active shape model
Image Vision Comput
(1997) - et al.
Automatic lumbar vertebral identification using surface-based registration
J Biomed Inform
(2001) - et al.
Using deformable model surface to segment 3-D images and infer differential structures
Compute Vision Graph Image Process
(1992) - et al.
Medical image analysis: progress over two decades and the challenges ahead
IEEE Trans Pattern Anal Machine Intell
(2000) - et al.
Snakes: active contour models
Int J Comput Vision
(1988) - et al.
Seeded region growing
IEEE Trans Pattern Anal Machine Intell
(1994) - et al.
Deformable boundary finding in medical images by integrating gradient and region information
IEEE Trans Med Imaging
(1996) - et al.
3D statistical shape models for medical image segmentation.
Automatic segmentation of lumbar vertebrae on digital radiographs using linked active appearance models
Proc Med Image Understanding Anal
Cited by (92)
Modeling the Correlation of Human Vertebral Body Volumes
2023, IFAC-PapersOnLineFully automated 3D segmentation and separation of multiple cervical vertebrae in CT images using a 2D convolutional neural network
2020, Computer Methods and Programs in BiomedicineCitation Excerpt :To address these problems, many studies have proposed and evaluated semi- or fully automated vertebral segmentation algorithms. Various strategies range from conventional image-processing methods, such as model-based methods [4], region growing [5], and the level set method [6] to recent deep learning-based methods [3]. Among them, the deep learning-based method shows performance superior to that of the conventional methods.
A deep learning framework for vertebral morphometry and Cobb angle measurement with external validation
2022, European Spine JournalFuture Perspective of Augmented Reality in Minimally Invasive Spine Surgery
2022, Technical Advances in Minimally Invasive Spine Surgery: Navigation, Robotics, Endoscopy, Augmented and Virtual Reality
Yiebin Kim is a Ph.D. candidate in School of Electronic Engineering at the Soongsil University, Seoul, Korea. He received his BS degree (in 2002) in Department of Electronic engineering and MS degree (in 2004) in Department of Information Telecommunication engineering from Soongsil University. His major research interests include medical image processing, computer vision, and robotics.
Dongsung Kim received his BS (in 1986) and MS (in 1988) in Department of Electronic engineering from Seoul National University and his Ph.D. (in 1994) in computer vision from University of Southern California. From 1994 to 1996, he was a postdoctoral researcher at the University of California, Riverside, working on image segmentation. Since 1996, he joined Soongsil University, Korea and is an associate professor in School of Electronic Engineering. His research interests include medical image processing and computer vision, in particular segmentation tools for commercial and research purposes.
- 1
Tel.: +82 2 822 8441; fax: +82 2 821 7653.