Identification of L5 Vertebra on Lumbar Spine Radiographs Using Deep Learning

doi:10.21203/rs.3.rs-3086867/v1

Download PDF

Research Article

Identification of L5 Vertebra on Lumbar Spine Radiographs Using Deep Learning

https://doi.org/10.21203/rs.3.rs-3086867/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background: Accurate and reliable spine numbering is critical for diagnosis, pre-procedural and preoperative planning, and treatment for spine pathology; however, it can be sometimes difficult to enumerate spine segment. Deep learning is an advanced machine-learning approach used in several medical fields. In this study, we aimed to develop a deep learning model using an object detection algorithm to identify the L5 vertebra on anteroposterior lumbar spine radiographs, and we assessed its detection accuracy.

Methods: A total of 150 participants for whom both anteroposterior whole spine and lumbar spine radiographs were available were retrospectively recruited. Anteroposterior lumbar spine radiographs of 150 patients were used as input data. Of the 150 images, 105 (70%) were randomly selected as the training set, and the remaining 45 (30%) were assigned to the validation set. YOLOv5x of the YOLOv5 family model was employed to detect the L5 vertebra area.

Results: The mean average precisions 0.5 and 0.75 of the trained L5 detection model were 99.2% and 96.9%, respectively. The model’s precision and recall were 95.7% and 97.8%, respectively. Of the validation data, 93.3% were detected correctly.

Conclusion: Our deep learning model showed an outstanding ability to identify L5 vertebrae for spine enumeration and numbering.

Deep learning

L5 vertebra

Enumeration

Deep learning

Algorithm

Spine radiograph

Accurate and reliable spine numbering is critical for pathological diagnosis and pre-procedural or preoperative planning. However, it can be difficult in patients with lumbosacral transitional vertebrae, especially when whole-spine images are not available [1]. There are two types of lumbosacral transitional vertebrae: sacralization of L5, a congenital spinal anomaly in which an elongated transverse process of the last lumbar vertebra fuses with the "first" sacral segment in varying degrees [1–3], and lumbarization of S1, a spinal anomaly in which the first and second segments of the sacrum are not fused, and consequently, the lumbar spine appears to have six rather than five vertebrae [1–4].

The lumbar vertebrae can be accurately numbered using whole-spine radiography or whole-spine magnetic resonance (MR) imaging by counting down from C2 to the sacrum [5, 6]. In cases without the availability of full-spine imaging, accurate numbering of the lumbar vertebrae is frequently confusing. Up to 32% of neurosurgeons have reported that they performed wrong-level spinal surgery at least once in their careers [7]. Therefore, prior to a spinal procedure or operation, clinicians should perform whole-spine imaging in patients with spine disorders to accurately identify the lumbar vertebral level.

Deep learning techniques have recently emerged as powerful methods for automatically learning feature representations from data [8–10]. In particular, these techniques significantly improve object detection [11]. Object detection determines whether there are any instances of objects from specified categories in an image and, if so, provides the spatial location and extent of each object instance via a bounding box [11]. Object detection, as the foundation of image comprehension and computer vision, serves as the basis for solving complex or high-level vision tasks such as segmentation, object tracking, image captioning, activity recognition, and event detection [12]. In the medical field, object detection is widely used to identify various pathologies such as malignancy, pneumonia, and bony fractures on medical images [13, 14]. We hypothesized that object detection can be applied for identifying the lumbar vertebral level.

In the current study, we developed a deep learning model using an object detection algorithm to identify the L5 vertebra on a lumbar spine radiograph and evaluated its accuracy.

Participants

A total of 150 participants aged ≥ 20 years who visited the spine center of a university hospital between June 2016 and December 2021 (mean age, 66.86 ± 12.97 years; male:female, 84:66) were retrospectively recruited for this study. Both anteroposterior whole-spine and lumbar-spine radiographs were stored in the hospital picture archiving and communication system for each included participant. The study protocol was approved by the institutional review board of Yeungnam university medical center (IRB no. 2023-02-005) and followed the Declaration of Helsinki. The requirement for written informed consent was waived because of the study’s retrospective nature and that was also approved by our institutional review board (YUMC IRB no. 2023-02-005).

Input data

Among the 150 images, 105 (70%) were randomly selected as the training set, whereas the remaining 45 (30%) were assigned to the validation set using the scikit-learn package. Table 1 lists the dataset details. On each patient's anteroposterior lumbar spine radiograph, the L5 vertebra was identified by counting down from C2 to the sacrum on an anteroposterior whole-spine radiograph.

Table 1

Dataset configuration
	Sample size (with metal implant)	Sample size (no metal implants) immmedfsssdfimimplants)implants)	Total
Training Set	97	8	105 (70%)
Validation Set	41	4	45 (30%)
Total	138	12	150 (100%)

To label the L5 vertebra, LabelImg (Heartex, San Francisco, CA, USA; https://github.com/heartexlabs/labelImg), an open-source software widely used to annotate and label bounding boxes on images, was adopted. On each patient's anteroposterior lumbar spine radiograph, the L5 vertebra was identified by counting down from C2 to the sacrum on the anteroposterior whole-spine radiograph. Rectangular bounding boxes were created to cover all L5 areas with a class named “L5.” All annotation files were exported in YOLO format.

Deep learning model development

Python 3.9.12 and Pytorch 1.12.0 were used to develop a deep learning model for detecting L5 vertebrae on lumbar spine radiographs. YOLOv5x of the YOLOv5 family model was employed to detect the L5 vertebra area. YOLOv5 is a family of state-of-the-art object-detection architectures pretrained on the COCO dataset. Our L5 detection model detected only one class of objects. Therefore, the YOLOv5x model with the highest accuracy was employed because accuracy is more important than detection speed [15].

We used a stochastic gradient descent optimizer and generalized intersection over union loss function to optimize our model. Data augmentation techniques, such as image flip, mixup, and mosaic, improved the model's generalization performance. Table 2 lists the details of the model’s experimental hyperparameters.

Table 2

Model hyperparameters for training
Parameters	Value	Remark
Optimizer	SGD
Loss function	GIOU
lr0	0.01	initial learning rate
lrf	0.1	final OneCycleLR learning rate (lr0 * lrf)
Weight_decay	0.0005	optimizer weight decay 5e-4
momentum	0.937	SGD momentum/Adam beta1
hsv_h	0.015	image HSV-Hue augmentation (fraction)
hsv_v	0.4	image HSV-Value augmentation (fraction)
fliplr	0.5	image flip left-right (probability)
mosaic	1.0	image mosaic (probability)
mixup	0.1	image mixup (probability)
SGD, stochastic gradient descent; GIOU, generalized intersection over union; lr, learning rate; lr0, initial learning rate; lrf, final learning rate; HSV, hue, saturation, value

Performance analysis

The mean average precisions (mAPs) 0.5 and 0.75 were used to measure the performance of the L5 detection model. The mAP_0.5 and mAP_0.75 metrics were considered to indicate “good detection” when the intersection over union, which is an indicator of the degree to which two areas overlap between the areas of the detected and real objects, was ≥ 50% and ≥ 75%, respectively.

Failure analysis was performed to provide additional information for understanding the causes of incorrectly detected cases and to help researchers better understand the strengths and weaknesses of the deep learning algorithm [16].

The mAP_0.5 and mAP_0.75 metrics of the trained L5 detection model were 99.2% and 96.9%, respectively, indicating outstanding performance of our L5 detection model. The precision and recall were 95.7% and 97.8%, respectively. Figure 1 shows the changes in mAP_0.5 and mAP_0.75 metrics by training epoch. Of the 45 validation data points, the L5 vertebra was correctly detected with ≥ 90% confidence in 42 (93.3%; Fig. 2).

Failure analysis was performed to aid our understanding of the incorrectly detected cases. The L5 vertebra was completely miss-detected in two (4.4%) of the 45 validation data points (Figs. 3A, 3B). The metal implants in the middle image seem to have caused the miss detection. Figure 3C shows an incorrectly detected case. In our L5 detection model, two vertebrae were selected as L5. The vertebra in the upper bounding box was presented as L5 with 87% confidence, whereas that in the lower bounding box was presented as L5 with 53% confidence; however, the vertebra in the lower bounding box was the actual L5 vertebra.

In the current study, we developed a deep learning model to detect the L5 vertebra using YOLOv5x. The mAP_0.5 metric of our developed model was 98.2%. The results of the failure analysis showed that 42 cases (93.3%) among the validation dataset were correctly identified, whereas only 3 (6.7%) were incorrectly detected. Considering that ≥ 90% accuracy in a deep learning classification model is outstanding, our L5 vertebra detection model seems to have outstanding performance [17].

We aimed to confirm the exact location of L5 using deep learning. The accurate determination of the lumbar spine segment is of great importance [18]. Lumbar spine pathology and related clinical symptoms are significantly associated with specific lumbar nerve roots, which can be localized to the lumbar spine level [19]. Thus, selecting improper lumbar spine levels makes identifying the wrong lumbar nerve root likely, resulting in the management of improper lumbar segment levels during interventional or surgical procedures [20]. Historically, several methods for accurately identifying the number of a lumbar spine level have been reported, including a method using the 12th rib, iliolumbar ligament, kidney level, and craniocaudal level on whole-spine radiographs. However, conventional methods are not accurate [21–23]; therefore, spine physicians have endeavored to develop methods of accurately identifying the number of a lumbar spine segment.

In particular, accurate selection of the L5 level is important in the clinical setting, as most lumbar pathologies requiring intervention or surgery are located at L4–5 and L5–S1 [24]. However, identifying the L5 vertebra using conventional methods is difficult for the following reasons. L5 is easily miscounted when sacralization or lumbarization occurs. In the case of spondylolisthesis, the lumbar vertebrae are easily numbered erroneously when only the anteroposterior radiograph is examined. Additionally, in cases of severe osteoporosis, spinal or pelvic deformity, or congenital spinal pathologies, an accurate L5 selection is more difficult than usual. When spinal computed tomography (CT) or MR imaging (MRI) is performed together with lumbar spinal radiography, the determination of the L5 vertebra is not quite difficult. However, in clinical practice, spinal radiography is typically performed without spinal CT or MRI. Therefore, during spinal surgery or intervention, the L5 vertebra could be incorrectly determined by clinicians.

Regarding the determination of the L5 vertebra, many studies have examined patients with lumbosacral transitional vertebrae [22, 23]. There are two main methods for determining the location of the L5 vertebra. The first method identifies the L5 vertebra on the basis of characteristic vertebrae or structures. By counting down the number of vertebral bodies from C2, clinicians can accurately identify L5. Moreover, identification of the T12 vertebra by checking the vertebra where the last rib is attached can reveal the L1 vertebra, and the fourth vertebra below L1 is the L5. The L5 vertebra can also be identified by its transverse process using the Ferguson view. However, in typical clinical situations, whole-spine radiography is not performed. The method used to initially identify T12 is less accurate owing to the dysplastic rib, and the method using the Ferguson view has a sensitivity of 76–84%. The second method involves determining the L5 vertebra by identifying the iliac crest tangent sign in the coronal planes of MRI scans [21–23]. However, this method has a sensitivity of 81%. Compared with these classical methods used to count the number of vertebrae, our developed deep learning model shows high accuracy (93.3%) for detecting the L5 vertebra on anteroposterior lumbar spine radiographs. A deep learning model is characterized by a multilayer structure with multiple hidden layers and provides a higher ability than a traditional shallow learning model [8]. We believe that our deep learning model extracts valuable features that differentiate the L5 vertebra from other vertebral levels.

Certain limitations should be pointed out in the present study. First, it included a relatively small amount of imaging data. Second, we used a dataset from a single hospital. Since each hospital may have different image acquisition methods and resolutions, it may be important to acquire multi-center data. Third, we did not compare the accuracy of our developed model with that of human labor. Fourth, only AP radiography was used as input data for deep learning model development, and it is thought that the accuracy could be further improved if additional radiograph (e.g., lateral radiograph) was included in input data. Finally, we could not find a way to improve the problems mentioned in failure analysis. To increase the accuracy and versatility of our deep learning model and measure its performance more accurately, further study with larger samples and enhanced study design would be necessary.

We developed a deep learning model for detecting the L5 vertebra using YOLOv5x that showed outstanding performance. We believe that our model may be useful in clinical practice. To our knowledge, our study offers the first deep learning model to detect the L5 vertebra. Further study with larger samples is necessary to better demonstrate the accuracy and performance of our deep learning model with YOLOv5x.

Ethics approval and consent to participate: The study protocol was approved by the institutional review board of Yeungnam university medical center (IRB no. 2023-02-005) and followed the Declaration of Helsinki. The requirement for written informed consent was waived because of the study’s retrospective nature and that was also approved by our institutional review board (YUMC IRB no. 2023-02-005).

Consent for publication: Not applicable.

Availability of data and materials: The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Competing interests: The authors declare no conflict of interest.

Funding: Not applicable.

Authors' contributions: Conceptualization, G.W.L.; methodology, G.W.L.; software, W.T.P.; validation, G.W.L.; formal analysis, M.C.C., G.W.L.; investigation, J.K.K., M.C.C., W.T.P., and G.W.L.; resources, W.T.P.; data curation, W.T.P.; writing—original draft preparation, J.K.K. and M.C.C.; writing—review and editing, W.T.P. and G.W.L.; visualization, G.W.L.; supervision, G.W.L.; project administration, G.W.L.; funding acquisition, G.W.L. All authors have read and agreed to the published version of the manuscript.

Acknowledgements: Not applicable.

Peckham ME, Hutchins TA, Stilwill SE, Mills MK, Morrissey BJ, Joiner EAR, Sanders RK, Stoddard GJ, Shah LM. Localizing the L5 Vertebra Using Nerve Morphology on MRI: An Accurate and Reliable Technique. AJNR Am J Neuroradiol. 2017;38:2008-2014.
Hughes RJ, Saifuddin A. Imaging of lumbosacral transitional vertebrae. Clinical Radiology. 2004;59:984–991.
Mahato NK. Redefining lumbosacral transitional vertebrae (LSTV) classification: integrating the full spectrum of morphological alterations in a biomechanical continuum. Medical Hypotheses. 2013;81:76–81.
Jancuska JM, Spivak JM, Bendo JA. A Review of Symptomatic Lumbosacral Transitional Vertebrae: Bertolotti's Syndrome. Int J Spine Surg. 2015;9:42.
Paik NC, Lim CS, Jang HS. Numeric and morphological verification of lumbosacral segments in 8280 consecutive patients. Spine (Phila Pa 1976). 2013;38:E573–578.
Paik NC, Lim CS, Jang HS. Numbering of vertebrae on MRI using a PACS cross-referencing tool. Acta Radiol. 2012;53:785–789.
Jhawar BS, Mitsis D, Duggal N. Wrong-sided and wrong-level neurosurgery: a national survey. J Neurosurg Spine. 2007;7:467–472.
Kim JK, Chang MC, Park D. Deep Learning Algorithm Trained on Brain Magnetic Resonance Images and Clinical Data to Predict Motor Outcomes of Patients With Corona Radiata Infarct. Front Neurosci. 2022;15:795553.
Kim JK, Choo YJ, Choi GS, Shin H, Chang MC, Park D. Deep Learning Analysis to Automatically Detect the Presence of Penetration or Aspiration in Videofluoroscopic Swallowing Study. J Korean Med Sci. 2022;37:e42.
Kim JK, Choo YJ, Shin H, Choi GS, Chang MC. Prediction of ambulatory outcome in patients with corona radiata infarction using deep learning. Sci Rep. 2021;11:7989.
Liu L, Ouyang W, Wang X, et al. Deep Learning for Generic Object Detection: A Survey. Int J Comput Vis. 2020;128:261–318.
Li Y, Ouyang W, Wang X, Tang X. Vip-CNN: Visual phrase guided convolutional neural network. Conference on Computer Vision and Pattern Recognition, 2017.
Bhandari A, Prasad PWC, Alsadoon A, Maag A. Object detection and recognition: using deep learning to assist the visually impaired. Disabil Rehabil Assist Technol. 2021;16:280-288.
Yang R, Yu Y. Artificial Convolutional Neural Network in Object Detection and Semantic Segmentation for Medical Imaging Analysis. Front Oncol. 2021;11:638182.
Liu, K., Tang, H., He, S., Yu, Q., Xiong, Y., & Wang, N. Performance validation of YOLO variants for object detection. In Proceedings of the 2021 International Conference on Bioinformatics and Intelligent Computing. pp. 239-243, 2021.
Mongan, J., Moy, L., & Kahn Jr, C. E. Checklist for artificial intelligence in medical imaging (CLAIM): a guide for authors and reviewers. Radiology. Artificial Intelligence, 2(2) (2020).
Mandrekar JN. Receiver operating characteristic curve in diagnostic test assessment. J Thorac Oncol. 2010;5:1315-1316.
Kim GU, Chang MC, Kim TU, Lee GW. Diagnostic Modality in Spine Disease: A Review. Asian Spine J. 2020;14:910-920.
Kasai Y, Paholpak P, Wisanuyotin T, Sukitthanakornkul N, Hanarwut P, Chaiyamoon A, Iamsaard S, Mizuno T. Incidence and Skeletal Features of Developmental Cervical and Lumbar Spinal Stenosis. Asian Spine J. 2022 May 10. doi: 10.31616/asj.2022.0015. Epub ahead of print. PMID: 35527532.
Kato S, Terada N, Niwa O, Yamada M. Factors Affecting Incomplete L5/S Posterior Lumbar Interbody Fusion, Including Spinopelvic Sagittal Parameters. Asian Spine J. 2022;16:526-533.
Tatara Y, Niimura T, Sekiya T, Mihara H. Changes in Lumbosacral Anatomy and Vertebral Numbering in Patients with Thoracolumbar and/or Lumbosacral Transitional Vertebrae. JBJS Open Access. 2021;6:e20.00167.
Tureli D, Ekinci G, Baltacioglu F. Is any landmark reliable in vertebral enumeration? A study of 3.0-Tesla lumbar MRI comparing skeletal, neural, and vascular markers. Clin Imaging. 2014;38:792-796.
Zhou S, Du L, Liu X, Wang Q, Zhao J, Lv Y, Yang H. Quantitative measurements at the lumbosacral junction are more reliable parameters for identifying and numbering lumbosacral transitional vertebrae. Eur Radiol. 2022;32:5650-5658.
Paik NC, Lim CS, Jang HS. Numbering of vertebrae on MRI using a PACS cross-referencing tool. Acta Radiol. 2012;53:785-789.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Identification of L5 Vertebra on Lumbar Spine Radiographs Using Deep Learning

Status:

Version 1

Abstract

Figures

Background

Methods

Participants

Input data

Deep learning model development

Performance analysis

Results

Discussion

Conclusions

Declarations

References

Additional Declarations

Status:

Version 1