Study of lumbar spine activity regularity based on Kanade-Lucas-Tomasi algorithm

https://doi.org/10.1016/j.bspc.2018.12.023Get rights and content

Highlights

  • Accuracy of the average errors achieved for displacement and angle of KLT - based algorithm were high.

  • Quantitative fluoroscopy (QF) can obtain continuous, dynamic image sequences of lumbar motion at low radiation doses.

  • More smoothness and stability in the intervertebral kinematics parameters were obtained by KLT - based algorithm.

Abstract

The diagnosis of lumbar spine pathologies such as lower back pain (LBP) is extremely difficult, and a consensus regarding a diagnostic standard has not yet been reached. Accurate measurement of intervertebral kinematic parameters helps to understand the etiology and diagnosis of related diseases, and to evaluate the subsequent treatment. Quantitative fluoroscopy (QF) which can obtain continuous, dynamic image sequences of lumbar motion at low radiation doses were introduced. The estimation of intervertebral kinematics depends on the precise localization of lumbar vertebrae. Usually, vertebral corners are identified and tracked by automated methods or manual method. This study presents a novel method for vertebral tracking based on the KLT (Kanade-Lucas-Tomasi, KLT) algorithm for intervertebral kinematics analysis. Accuracy of the proposed method was tested using a sagittal human image which contained two vertebral bodies, and the average errors achieved for displacement and angle were within 0.25 mm and 0.3°, respectively. To test the proposed method under clinical conditions, totally 11 volunteers were recruited and a result of more smoothness and stability in the intervertebral kinematics was obtained with comparing to the manual method by an experienced and trained clinician. In conclusion, the proposed method is an effective, accurate way to assess lumbar vertebrae tracking.

Introduction

Lumbar diseases, such as lower back pain (LBP), are common chronic diseases in middle-aged and older individuals, and degenerative lumbar instability is considered the most important factor causing LBP, sciatica and other symptoms [1,2]. Due to the complexity of lumbar movement and the lack of specificity in clinical manifestations, many studies have focused on the relationship between lumbar instability and clinical symptoms, although the pathogenesis, clinical diagnosis and diagnostic criteria associated with imaging parameters of lumbar spine instability are still debated. For degenerative lumbar instability, a standardized diagnostic criterion has not yet been established. Degenerative lumbar spine disease can lead to increased segmental lumbar motion [3,4], and abnormal intervertebral motion may cause back pain [5]. These reasons confirm the importance of measuring intervertebral kinematic parameters. Since measuring the parameters directly in vivo is difficult, we employed radiographic lumbar images in this study to track lumbar activity, intending to collect corresponding dynamic lumbar spine parameters. At present, in clinical practice, the hyperextension and hyperflexion displacements and rotation angles in the sagittal plane on an X-ray are typically used as criteria for determining whether lumbar spine instability is present. However, lateral flexion, rotation and other activities are not assessed in the lumbar spine. Moreover, lumbar activity is a complex, three-dimensional movement, and static hyperextension and hyperflexion do not reflect its dynamic complexity. Furthermore, clinically common pain reactions caused by lumbar instability often occur during lateral flexion, rotation, or slight flexion and extension movements rather than in a static state. Traditional X-ray transmission technology has drawbacks, such as large radiation doses and challenging continuous image collection [6]. Quantitative fluoroscopy (QF) overcomes these drawbacks, obtaining continuous dynamic image sequences of moving objects using a low dose [7,8]. Our study utilizes QF to acquire dynamic medical images that were processed to achieve vertebral body position tracking, calculate relevant dynamic parameters, and advance the study of vertebral body motion patterns using QF lumbar images.

The key point when tracking lumbar vertebral bodies is vertebral body registration, which has attracted significant attention in recent years. In 1989, Breen et al. [9] proposed a manual method in order to calculate the relevant parameters. This method is labor-intensive and requires a skilled operator. Muggleton et al. [10], Bifulco et al. [11] and Penning et al. [12] utilized a template matching approach and relied on a cross-correlation function for lumbar vertebral body tracking. Although this method is simple and easy to implement, it is sensitive to geometric distortion caused by rigid templates. The template matching algorithm is also a global match in the tracking region; therefore, its computational complexity is high. In 2006, Wong et al. [13] suggested a learning-based method to perform lumbar motion tracking studies. Firstly, a support vector machine (SVM) was used to learn the texture modeled by Markov Random Fields (MRF). Then, the edge was detected using the texture information along the snake, and the tracking was performed after a Kalman filter. This approach had several drawbacks, including the computational complexity of long fluoroscopic sequences and sensitivity to illumination and contrast nonhomogeneities, which was caused by dependence on the modeled texture pattern intensity. McDonald et al. [14] and Lin et al. [15] applied the analysis to the lumbar vertebral body in three-dimensions. However, the greatest clinical interest was in flexion and extension motions in the sagittal plan, not the overall multidimensional structure. Balkovec et al. [16] proposed a straightforward iterative template matching method; unfortunately, this method is time consuming, and required the establishment and constant updating of template.

This study utilizes a feature-based KLT (Kanade-Lucas-Tomasi, KLT) algorithm [17,18] that is prevalent in visual tracking studies and is derived from the optical flow. The algorithm is based on feature points, which are tracked by utilizing affine transformation, and does not require global matching like the template matching algorithm. The KLT is commonly used for agriculture, visual monitoring, image compression, image splicing and 3D reconstruction, among others. Although the KLT algorithm has been applied to 2D/3D thorax image fusions [19] and swallowing studies [20], there are few studies examined its application to vertebral body tracking. In this work, a faster and more accurate method based on KLT for sagittal lumbar vertebrae tracking is presented (Fig. 1). By introducing KLT, the results of lumbar vertebra tracking can reach a high level of accuracy as well as with a low cost time. In addition, the Harris feature point detection method that could adapt for rotation transformations very well was introduced. A sagittal human image containing two vertebrae as a simulation model was used to assess the proposed method. And totally 11 volunteers’ fluoroscopic sequences were estimated using the proposed method with a comparison to the manual method by an experienced and trained clinician because of the unrealistic to measure ground truth in vivo.

Section snippets

Image acquisition

All images were taken with an electric, digital, 3D C-arm X-ray provided by the Wuxi Affiliated Hospital of Nanjing University of Chinese Medicine. During measurement, the digital frame grabber acquired images at a rate of 15 frames per second (fps). The image size was 1024×1024 pixels (0.29 mm by 0.29 mm per pixel). To obtain better tracking results during actual processing, we applied inter-frame processing methods to reduce the influence of the background. The processed video sequence was

Results

Fig. 5 shows an original fluoroscopic lumbar image (Fig. 5(a)) and an edge-enhanced image after edge preserving filtering (Fig. 5(b)) which the background soft tissue was smoothed. Fig. 6 shows a comparison of tracking results before and after edge-enhance. Here take the 150th frame as an example to illustrate the edge-enhanced effectiveness. Fig. 6 (a) shows the initialized frame and Fig. 6 (b) shows the tracking results of the 150th frame with edge-enhance, respectively. To observe the

Discussion and conclusion

In clinical practice, fluoroscopy images of lumbar vertebrae are usually complicated by pincushion distortion, soft tissue movement, blurred edges, low image contrast, brightness change, and noise influence. From Fig. 6 (b) and (c), we found that the edge-enhanced images had better tracking results than without edge-enhance. The template matching algorithm and learning-based method were sensitive to illumination and contrast nonhomogeneities. However, from the tracking results of the original

Acknowledgements

This work was supported by National Natural Science Foundation of China under Grant 81473693, Natural Science Foundation of Fujian Province, China under Grant 2017J01116 and Promotion Program for Young and Middle-aged Teacher in Science and Technology Research of Huaqiao University under Grant Z16J0070.

References (24)

  • A. Ahmadi et al.

    Kinematic analysis of dynamic lumbar motion in patients with lumbar segmental instability using digital videofluoroscopy

    Eur. Spine J.

    (2009)
  • A.C. Breen et al.

    Measurement of Intervertebral motion using quantitative fluoroscopy: Report of an International forum and proposal for use in the assessment of degenerative disc disease in the lumbar spine

    Adv. Orthop.

    (2012)
  • Cited by (0)

    View full text