BUU-LSPINE: A Thai Open Lumbar Spine Dataset for Spondylolisthesis Detection

Abstract

(1) Background: Spondylolisthesis, a common disease among older individuals, involves the displacement of vertebrae. The condition may gradually manifest with age, allowing for potential prevention by the research of predictive algorithms. However, one key issue that hinders research in spondylolisthesis prediction algorithms is the need for publicly available spondylolisthesis datasets. (2) Purpose: This paper introduces BUU-LSPINE, a new dataset for the lumbar spine. It includes 3600 patients’ plain film images annotated with vertebral position, spondylolisthesis diagnosis, and lumbosacral transitional vertebrae (LSTV) ground truth. (4) Methods: We established an annotation pipeline to create the BUU-SPINE dataset and evaluated it in three experiments as follows: (1) lumbar vertebrae detection, (2) vertebral corner points extraction, and (3) spondylolisthesis prediction. (5) Results: Lumbar vertebrae detection achieved the highest precision rates of 81.93% on the AP view and 83.45% on the LA view using YOLOv5; vertebral corner point extraction achieved the lowest average error distance of 4.63 mm on the AP view using ResNet152V2 and 4.91 mm on the LA view using DenseNet201. Spondylolisthesis prediction reached the highest accuracy of 95.14% on the AP view and 92.26% on the LA view of a testing set using Support Vector Machine (SVM). (6) Discussions: The results of the three experiments highlight the potential of BUU-LSPINE in developing and evaluating algorithms for lumbar vertebrae detection and spondylolisthesis prediction. These steps are crucial in advancing the creation of a clinical decision support system (CDSS). Additionally, the findings demonstrate the impact of Lumbosacral transitional vertebrae (LSTV) conditions on lumbar detection algorithms.

1. Introduction

1.1. Background Knowledge

Spondylolisthesis occurs when a vertebra slips out of its position relative to the vertebra below it. The disease is common among elderly individuals [1], leading to chronic lower back pain, nerve compression, and even paralysis in severe cases [2]. Therefore, predicting spondylolisthesis is crucial for successful treatment and preventing complications. Medical imaging, specifically X-ray imaging, is widely used for detecting spondylolisthesis [3,4]. However, manual detection of spondylolisthesis from X-ray images can be time-consuming, prone to errors, and dependent on the radiologist’s skill [5].

Several automated methods have been proposed for spondylolisthesis detection to address these limitations [6,7,8,9]. However, these methods often rely on private datasets, which limit repeatability and direct accuracy comparison. This paper introduces a new dataset of lumbar spine X-ray images called BUU-LSPINE, which includes ground truth annotations of the vertebra position and spondylolisthesis diagnosis. Our dataset is valuable because it has many high-quality X-ray images with diverse demographics and types of spondylolistheses. Additionally, our dataset includes crucial clinical information such as patient age and gender, the presence of lumbosacral transitional vertebrae (LSTVs) [10], and pixel spacing [11], making it a unique resource for developing and evaluating automated spondylolisthesis detection algorithms.

LSTV, also called Bertolotti’s syndrome [12], is a congenital disability that affects the last lumbar vertebra [13,14,15]. Generally, the human lumbar spine consists of five vertebrae (L1 to L5), but patients with LSTV have a different number of lumbar vertebrae [16]; patients with four lumbar are denoted as having a sacralization of L5, and patients with six lumbar are denoted as having a lumbarization of S1. Misclassification of LSTV can affect the diagnosis and medical treatment [17] and affect the results of algorithms that use lumbar vertebrae. For example, Deng et al. [18] indicated some failed prediction results from spinal disease due to lumbarization and sacralization.

Pixel spacing in medical images specifies the distance, in millimeters (mm), between the centers of adjacent pixels along the x and y axes of the picture. This attribute is critical in converting distance in pixels on an image to physical space [19].

1.2. Literature Review

Most studies on spondylolisthesis rely on a private dataset, as seen in studies 1 to 6 in

Table 1. Literature review for spinal detection algorithms.

No	Study	Source	Spondylolisthesis Diagnosis	Vertebra Position	LSTV Info	Number of Patients	View(s)
1.	Zhao et al. [6]	Private	✓	✓	✕	150	LA
2.	Saravagi et al. [21]	Private	✓	✕	✕	337	AP, LA
3.	Liao et al. [22]	Private	✓	✓	✕	558	LA
4.	Varçın et al. [8]	Private	✓	✓	✕	600	LA
5.	Trinh et al. [23]	Private	✓	✓	✕	706	LA
6.	Nguyen et al. [7]	Private	✓	✓	✕	1000	LA
7.	Cai et al. [20]	Public	✓ (synthesized)	✕	✕	47	LA
8.	Fraiwan et al. [9]	Public	✓	✕	✕	338	AP
9.	Chu et al. [24]	Public	✕	✓	✕	23	3D
10.	Sekuboyina et al. [25]	Public	✕	✓	✕	355	3D
11.	Masood et al. [26]	Public	✕	✓	✕	514	LA
12.	Deng et al. [18]	Public	✕	✓	✕	1005	3D
13.	Our proposed dataset	Public	✓	✓	✓	3600	AP, LA

. Unfortunately, this approach makes comparing their results to others difficult. Cai et al. [20] researched spondylolisthesis using a publicly available dataset. However, they had to synthesize the condition by manually displacing the vertebrae in X-ray images, as shown in Study 7 in Table 1. The only study with spondylolisthesis that publicly provides its dataset is from Fraiwan et al. [9], as shown in study 8 in Table 1. However, this dataset does not include vertebra position, which is crucial information to locate the disease position and its severity. Moreover, some cases have various spondylolisthesis locations in one patient, so identifying disease locations is essential to determine multiple instances of spondylolisthesis in one image.

While current public datasets for the human spine are great for detecting vertebrae, they do not include any ground truth for spinal diseases, like spondylolisthesis, as shown in studies 9 to 12 in Table 1. In contrast, our BUU-LSPINE dataset offers a comprehensive, diverse, and suitable dataset for spondylolisthesis detection. It includes both vertebral position and spondylolisthesis diagnosis, as shown in study number 13 in Table 1. Furthermore, our dataset comprises more patients than the other studies, as shown in the column number of patients. Moreover, including LA view images in our dataset is crucial in spondylolisthesis studies. This significance is highlighted by the presence of LA view images in numerous spondylolisthesis studies listed in Table 1. The LA view is vital for spondylolisthesis studies due to the potential consequences of vertebral slippage on the sagittal plane, such as spinal canal compression and the risk of nerve damage [4].

Our literature review identified a lack of publicly available datasets that include accurate vertebral positioning and a reliable spondylolisthesis ground truth. To address this gap, we created and published a public dataset with these ground truths, described in the next section.

2. Materials and Methods

2.1. BUU-LSPINE

BUU-LSPINE was developed by cooperation between Burapha University (BUU), Thailand, and the Korea Institute of Oriental Medicine (KIOM), Republic of Korea.

This study was conducted by the Declaration of BUU and approved by the Institutional Review Board of the BUU with the reference number HS024/2565(E3). The dataset collection underwent an anonymization process to eliminate all personal data. We preserved only the patients’ gender and age. All records in our dataset were assigned the four digits’ unique identification number. This information allows users access to each record anonymously. However, it is essential to note that some unique identification numbers are missing due to the filtering process.

BUU-LSPINE is a collection of PF X-ray images of 3600 patients, as shown in

Table 2. BUU-LSPINE description.

No	Field Name	Description
1.	Image Type	X-ray (PF)
2.	Image View	AP and LA
3.	Body Part	Lumbar spine (LSPINE)
4.	Total number of patients	3600 patients
5.	Total number of images	7200 images
6.	Total number of spondylolisthesis patients	621 patients
7.	Total number of spondylolisthesis instances	788 instances (Some patients include multiple instances)
8.	Datasets storage size	18.5 GB
9.	Ground truth	Vertebral position,
		Spondylolisthesis diagnosis,
		LSTV
10.	Lumbar vertebral position targets	The coordinate corner points of L1 to L5
11.	Spondylolisthesis diagnosis targets	Left laterolisthesis,
		Right laterolisthesis
		Anterolisthesis
		Retrolisthesis
12.	LSTV targets	Sacralization,
		Lumbarization
13.	File types	Images: .jpg
		Ground truth: .csv
14.	Locations	Burapha University Hospital, Thailand
15.	Year period	2000–2021
16.	Age range	6–97 years old
17.	Image dimensions	Height: 1434–3408 pixels, Width: 860–3040 pixels, Pixel Spacing: 0.125–0.175 mm
18.	X-ray imaging parameters	AP	kV: 70–75 kV mAs: 32–40 mAs
		LA	kV: 80–90 kV mAs: 50–60 mAs

. Each patient’s record has two views: an anteroposterior (AP) and a lateral (LA), as shown in Figure 1. Our dataset of 3600 patients comprises 7200 images because each has two views. All lateral images were flipped to face the left side. A record of our dataset includes two JPG X-ray images and two CSV ground truth files. Our dataset consists of three ground truths: vertebral position, spondylolisthesis, and LSTV. Further elaboration on the ground truths is provided in the following section.

The X-ray imaging process involves two essential parameters: the x-ray tube voltage (kV) and the x-ray tube current and the exposure time (mAs). Both parameters play a vital role in controlling the quality and contrast of the X-ray image, which depends on the thickness of the patients. Thicker patients require higher kV values to enhance the penetration capability of the X-ray radiation, while thinner patients necessitate lower kV values for optimal imaging results. For our plain film examinations, we typically used a range of 70–75 kV and 32–40 mAs for the AP view, while for the LA view, we utilized a range of 80–90 kV and 50–60 mAs, as shown in row 18 of Table 2.

2.2. Ground Truth Data

After we obtained permission from the Burapha University Hospital Institutional Review Board (IRB), we followed an annotation pipeline to make our dataset, as shown in Figure 2. Starting with collecting patient DICOM images from the hospital database, we filtered data using the following exclusion criteria to ensure that data consisted of all lumbar vertebrae and had good-quality pictures. Data with these criteria were excluded from our dataset.

Exclusion criteria:

• Low quality and high noise (challenging to distinguish the vertebrae);
• Embedded medical devices (cover the lumbar vertebrae area);
• Missing some of the lumbar vertebrae;
• Highly damaged vertebrae (challenging to determine the vertebral body).

We then submitted the filtered dataset to experts and physicians for review. Experts were computer vision experts, and physicians were radiologists and orthopedics doctors of KIOM and BUU. Next, we horizontally flipped all right-facing LA views to appear as left-facing views in the images. Subsequently, physicians annotated the ground truth of vertebral positions, LSTV, and made a diagnosis of spondylolisthesis. Ground truths are annotated on the CSV files, as shown in

Table 3. A ground truth CSV file in BUU-LSPINE (AP view).

Line	x1	y1	x2	y2	Class
L1a	876	167	1111	169	✕ 1
L1b	866	314	1120	314	0
L2a	871	344	1117	351	✕ 1
L2b	848	498	1120	511	0
L3a	842	553	1109	557	✕ 1
L3b	825	706	1118	725	0
L4a	826	775	1112	787	✕ 1
L4b	801	937	1114	958	0
L5a	798	983	1115	980	✕ 1
L5b	767	1164	1129	1158	0

¹ ✕ mean does not use. It is marked as ‘0’ in ground truth files but has no meaning.

for the AP view and

Table 4. A ground truth CSV file in BUU-LSPINE (LA view).

Line	x1	y1	x2	y2	Class
L1a	609	353	827	377	✕ 1
L1b	581	529	802	577	0
L2a	570	568	805	607	✕ 1
L2b	527	743	752	805	0
L3a	513	785	744	830	✕ 1
L3b	467	982	696	1025	0
L4a	450	1057	693	1063	✕ 1
L4b	430	1239	648	1259	0
L5a	427	1334	667	1295	✕ 1
L5b	452	1523	670	1474	0
S1a	483	1624	682	1508	0 2

¹ ✕ mean does not use. It is marked as ‘0’ in ground truth files but has no meaning. ² This mark is not spondylolisthesis but LSTV ground truth.

for the LA view. Finally, we conducted a dataset review meeting, inviting experts from various fields, such as orthopedics, radiology, and image processing. The complete version of the verified dataset, including X-ray images of the lumbar spine and the corresponding ground truth annotations of lumbar vertebrae position and spondylolisthesis diagnosis, is included.

2.2.1. Vertebral Positions

Ground truths of vertebral positions are the coordinates of all corner points of lumbar vertebrae. Two left and right corner points are combined as one vertebral edge line. Each row in Table 3 represents one vertebral edge line. There are two types of vertebral edge lines, the upper and lower edge line. The upper edge line ends with the ‘a’ character, and the lower edge line ends with the ’b’ character, as shown in Table 3.

In the AP view, there are ten vertebral edge lines, the upper and lower edges of five vertebrae, as shown on the left-hand side of Figure 3. Coordinates of all corner points are shown in CSV files with the same name as the corresponding images. Each row in a CSV file from Table 3 represents the coordinates of the left and right points defining a line segment, as shown on the right-hand side of Figure 3.

In the LA view, there are eleven vertebral edge lines. The first ten lines are the same as in the AP view. The extra line is annotated on the upper edge of the first sacral vertebra, as shown in the last row of Table 4 and Figure 4. The purpose of the S1a line is to be a reference line to measure the slippage of the L5 vertebra. This line allowed us to diagnose spondylolisthesis on the L5 vertebra, the most common vertebra that occurs in the disease. Moreover, the class of the S1a row is not a diagnosis of spondylolisthesis but the LSTV class of the patients. The LSTV ground truth details are described in Section 2.2.3.

2.2.2. Spondylolisthesis

Physicians investigated all vertebrae and diagnosed spondylolisthesis for each vertebra. The diagnosis results are marked as class 0 to 4 in the lower edge of the vertebra’s row and the last column of the CSV files. The total number of each class is shown in

Table 5. Total number of spondylolisthesis classes (number of vertebrae).

Class	Diagnosis	Female	Male	Total	View
0	Normal	22,276	12,936	35,212	AP, LA
1	Left laterolisthesis	44	17	61	AP
2	Right laterolisthesis	43	26	69	AP
3	Anterolisthesis	343	121	464	LA
4	Retrolisthesis	94	100	194	LA

. We must inform you that we marked spondylolisthesis on the lower edge line of each vertebra. All class labels on the upper edge line have no meaning.

2.2.3. LSTV

Diagnosing LSTV requires physicians to identify the L5 vertebra precisely. In sacralization instances, the L5 vertebra is fused with the sacrum and acts like a sacral vertebra. In lumbarization cases, the first sacral vertebra is separate from the sacrum and acts like the last lumbar. Moreover, lumbarization can be partially or entirely detached, requiring physicians’ skills to determine.

LSTV ground truths of BUU-LSPINE are marked on an S1a row in the CSV file of the LA view, as shown in the last row of Table 4. The total number of LSTV classes is shown in

Table 6. Total number of LSTV classes (number of patients).

Class	Diagnosis	Female	Male	Total
0	Normal	2181	1275	3456
20	Lumbarization	6	9	15
21	Sacralization	93	36	129

.

2.3. Benefit of BUU-LSPINE

BUU-LSPINE has different benefits among stakeholders. This dataset provides a valuable resource for researchers to research localization and identification of the human lumbar spine and spondylolisthesis. We demonstrated the usage of our dataset ground truth by conducting three experiments, which is explained in Section 2.4. Physicians require a system to support their work. The diagnostic process sometimes includes measuring and calculating the angle or distance between vertebrae, which is time-consuming. Our BUU-LSPINE datasets can be used to develop algorithms to help such tasks, as demonstrated in Section 2.4. Patients need reliable results. The BUU-LSPINE has 3600 records of different patients. It is large enough to be a reasonably representative sample for statistical analysis to help physicians make diagnostic decisions, as shown in Figure 5. Further research is required to inspect the full benefits of our dataset.

2.4. Experiments

We demonstrated the utility of BUU-LSPINE by conducting three different experiments: (1) detection of lumbar vertebrae; (2) extraction of corner points; (3) prediction of spondylolisthesis. In our tests, we used a Windows PC with an i7-8700 CPU running at 3.20 GHz, 16 GB of RAM, and an RTX 3080 graphics card. With the implementation of TensorFlow 2.7.0, we used Python 3.8.8.

2.4.1. Lumbar Vertebrae Detection

The targets of lumbar vertebrae detection experiment were detecting bounding boxes around vertebral bodies. The processes of this experiment are shown in Figure 6. The input was raw X-ray images, as shown in Figure 7a. We separated BUU-LSPINE into training and testing sets with a 0.8:0.2 ratio. The bounding boxes’ ground truth of the detection was created using corner points’ ground truth, as shown in Figure 7b. Then, we trained YOLOv5, MobileNetV1, MobileNetV2, ResNet50V1, ResNet101V1, ResNet152V1, ResNet50V2, ResNet101V2, ResNet152V2, EfficientDet D0, and EfficientDet D1 models, with a batch size of 4, a learning rate of 0.04, and all images resized to 640 × 640 pixels. The targets were bounding boxes of five vertebrae (‘APL1’, ‘APL2’, ‘APL3’, ‘APL4’, and ‘APL5’) on AP view and (‘LAL1’, ‘LAL2’, ‘LAL3’, ‘LAL4’, and ‘LAL5’) on LA view. Since the human spine typically consists of five lumbar vertebrae, our model chose only one candidate with the highest likelihood from each class. We stopped our training process early when there was no further improvement in precision. The detection results are shown in Figure 7c.

2.4.2. Vertebral Corner Points Extraction

The targets of the vertebral corner points extraction experiment were coordinates of all corner points of vertebral bodies. The processes of this experiment are shown in Figure 8. The input was raw X-ray images, as shown in Figure 9a. We separated BUU-LSPINE into training and testing sets with a 0.8:0.2 ratio. Pixel spacing ratios were used to convert corner points in pixels to millimeters. We then used it as the ground truth for training key point detection models, as shown in Figure 9b. Next, we chose MobileNetV1, MobileNetV2, ResNet50V1, ResNet101V1, ResNet152V1, ResNet50V2, ResNet101V2, ResNet152V2, DenseNet201, EfficientNetB0, and EfficientNetB1models, with a batch size of 4, a learning rate of 0.001, and all images resized to 640 × 640 pixels. The targets were 20 corner point coordinates for five vertebrae. We stopped our training process early when there was no further improvement in accuracy. A comparison between the ground truth corner points and the marked corner points is shown in Figure 9. The outcomes of the corner point extraction were four corner points of each vertebra, a total of five vertebrae, as shown in Figure 9c.

2.4.3. Spondylolisthesis Prediction

Tracking the progression of vertebrae displacement using the Meyerding classification allows for evaluating the severity of spondylolisthesis [27]. The Meyerding classification consists of five classes depending on the percentage of the displacement distance of the upper vertebra relative to its lower adjacent vertebra, as shown in Figure 10 and

Table 7. Classes of the Meyerding classification system.

Meyerding Classification	Percentage of Slip
Grade I	0–25%
Grade II	25–50%
Grade III	50–75%
Grade IV	75–100%
Grade V	>100%

. Analyzing the positions of pairs of vertebrae provides clues regarding symptoms of spondylolisthesis.

The direction of the slip determines the different types of spondylolistheses, as shown in Figure 11.

Treatments for spondylolisthesis include selective nerve blocks, which are aimed at reducing back pain, and posterior spinal fusion treatments that help stabilize the spinal structure, as shown in Figure 12.

The diagnosis of spondylolisthesis is based on the position of an adjacent pair of vertebrae, and the direction of the slip can identify its types. Therefore, our proposed method involved pairing every adjacent lumbar vertebra to begin the process, as shown in Figure 13. Then, we changed the reference coordinates of all corner points to the position of the left ends of the upper edge of the lower vertebra in an instance pair, as shown in Figure 14a. Through this analysis, we can establish the relationship between the positions of the upper and lower vertebrae in each pair of lumbar vertebrae. Lastly, we used classifiers to predict the spondylolisthesis in each pair, as shown in Figure 14b,c.

We conducted experiments to predict spondylolisthesis by creating a new dataset using BUU-LSPINE, as shown in Figure 13. We coupled the corner points of the upper vertebra with the lower vertebra as one record. Then, we used the upper-left corner point of the lower vertebra as a reference point for every point in the record, as shown in Figure 14a. The target ground truth of this experiment was the spondylolisthesis diagnosis ground truth from our proposed dataset, as shown in Figure 14b,c. Our new dataset consisted of 14,400 records in AP view (4 pairs of vertebrae: (L1–L2), (L2–L3), (L3–L4), and (L4–L5)) per image) and 18,000 records in LA view (5 pairs of vertebrae: (L1–L2), (L2–L3), (L3–L4), (L4–L5), and (L4–S1) per image). Since the ground truth diagnosis in our new dataset was highly imbalanced, we used under-sampling to match the minimum number of records in each class (61 of left laterolisthesis in the AP view and 194 of retrolisthesis in the LA view). Then, we split the dataset into training and testing sets with a ratio of 0.8:0.2, respectively.

3. Results

3.1. Lumbar Vertebrae Detection

We evaluated lumbar vertebrae detection results by using the mean average precision (mAP) metric over various intersection over union (IoU) thresholds ranging from 0.5 to 0.9 ([email protected]:0.95).

Table 8. Evaluation of models for detection of lumbar vertebrae.

Model	View	Mean Average Precision (%)			Recall
		@IoU [0.5]	@IoU [0.75]	@IoU [0.5, 0.95]
YOLOv5	AP	96.68	95.86	81.93	95.82
	LA	95.83	95.50	83.45	95.72
MobileNetV1	AP	90.70	88.57	75.77	80.81
	LA	89.48	88.72	76.02	81.91
MobileNetV2	AP	93.26	89.95	77.63	82.19
	LA	91.79	90.51	78.38	83.58
ResNet50V1	AP	92.22	90.08	77.70	81.43
	LA	91.03	89.95	77.73	81.97
ResNet101V1	AP	89.26	86.42	71.80	76.68
	LA	91.43	90.37	78.45	83.20
ReseNet152V1	AP	93.07	90.25	77.32	81.31
	LA	92.76	91.55	78.53	82.80
EfficientDet D0	AP	93.69	90.46	75.86	79.90
	LA	92.39	89.76	72.60	77.25
EfficientDet D1	AP	93.00	87.38	70.88	75.63
	LA	92.79	91.11	76.08	79.92

shows the evaluation of models. YOLOv5 achieved the highest precision of 81.93% and 83.45%, as well as the highest recall of 95.82% and 95.72% on the AP and LA views, respectively.

LSTV conditions can impact the detection of lumbar vertebrae since our model was trained under the assumption that every person’s spine consists of five lumbar vertebrae, as shown in Figure 15 and Figure 16. Patients with lumbarization, characterized by the presence of six vertebrae instead of the usual five, can lead to a consistent omission of one vertebra in our model’s detection, as shown in Figure 15b and Figure 16b. Patients with sacralization, where there are four vertebrae instead of the usual five, cause our model to attempt to locate the missing vertebra and mistakenly mark the position on the first sacrum instead of the fifth lumbar, as shown in Figure 15c and Figure 16c. These results highlight the importance of LSTV information in the lumbar spine dataset, which is included in our BUU-LSPINE dataset.

3.2. Vertebral Corner Point Extraction

We evaluated vertebral corner point extraction results using the Euclidean distance between the ground truth and the detected points.

Table 9. Evaluation of models for extraction of corner points.

Models	Average Error (Millimeter)		Average Error (% of Vertebra Width)
	AP	LA	AP	LA
MobileNetV1	5.10	5.56	9.72%ก	13.71%
MobileNetV2	6.78	7.29	12.92%	17.98%
ResNet50V1	5.59	6.37	10.65%	15.71%
ResNet101V1	9.81	8.51	18.70%	20.99%
ResNet152V1	5.85	5.86	11.15%	14.45%
ResNet50V2	5.27	5.83	10.05%	14.38%
ResNet101V2	5.48	6.15	10.45%	15.17%
ResNet152V2	4.63	5.41	8.83%	13.34%
DenseNet201	4.77	4.91	9.09%	12.11%
EfficientNetB0	143.80	240.38	274.17%	592.94%
EfficientNetB1	26.67	71.82	50.85%	177.16%

shows the evaluation results for models. Resnet152V2 achieved the lowest average error distance of 4.63 mm on the AP view and DenseNet201 achieved 4.91 mm on the LA view. According to statistics in our dataset, the average width of lumbar vertebrae in the AP view was 52.45 mm, which means the average error in our detection was 8.83%. The average width of lumbar vertebrae in the LA view was 40.54 mm, which means the average error in our detection was 12.11%. When comparing the average error in the percentage of vertebra width with the Meyerding classification system from Table 7, where each grade range is 25%, we observed relatively high average errors of 8.83% and 12.11%, respectively. This suggests that further improvements are needed in our feature point extraction models to achieve higher accuracy. Addressing this challenge remains an essential area of focus for researchers aiming to enhance the system’s performance.

The vertebral corner points provide valuable information for algorithms to predict spondylolisthesis, as shown in Figure 17 and Figure 18. We can measure the extent of vertebral slip by analyzing the relative positions of the upper with the adjacent lower vertebrae. Moreover, the direction of the slip can be crucial information in classifying the types of spondylolistheses.

3.3. Spondylolisthesis Prediction

We evaluated the spondylolisthesis prediction experiments by using three classifiers. The accuracies of the prediction are shown in

Table 10. Accuracies of spondylolisthesis prediction.

Classifier	View	Accuracy (%)
		Training Set	Testing Set
Decision Tree	AP	100.00%	92.77%
	LA	100.00%	77.38%
XGBoost	AP	100.00%	94.87%
	LA	100.00%	86.41%
SVM	AP	99.95%	95.14%
	LA	96.48%	92.26%

. SVM achieved the highest accuracy in the testing set of 95.14% in the AP view and 92.26% in the LA view. The prediction results are classes of pairs of adjacent vertebrae. The spondylolisthesis class indicates the slippage of the upper vertebra in each pair, as shown in Figure 19.

4. Discussion

We demonstrated the utilization of the BUU-LSPINE dataset in three experiments.

The first experiment uses PF images and corner points to create the bounding box ground truth to train lumbar vertebra detection models. The results demonstrate a high precision in detecting lumbar vertebrae positions. However, the experiment also reveals the influence of LSTV conditions on vertebrae detection algorithms, underscoring the importance of considering these conditions during the research of such algorithms.

The second experiment shows the utilization of the BUU-LSPINE’s corner point ground truth. The results show the potential of our datasets in researching lumbar vertebral corner point detection algorithms. However, the detected corner points from the best model in our experiment have a relatively high average error distance compared with the Meyerding Classification System. These errors may lead to incorrect classification of the grade of the vertebrae slippage.

In the third experiment, we introduce algorithms for predicting spondylolisthesis by pairing the adjacent vertebrae and adjusting the reference points of all corner points within the vertebrae pair. Then, we use classifiers to predict the spondylolisthesis based on the relation between all corner points and new reference points in each pair. The results show the performance in predicting the diseases. However, it is crucial to note that the capability of spondylolisthesis prediction depends on the performance of the vertebral corner point detection algorithms.

These experiments demonstrate the potential of the BUU-LSPINE dataset, but further research is required to explore our dataset capabilities. However, these experiments serve as an initial demonstration. Our dataset will catalyze new ideas and encourage others to use it in improving models, evaluating algorithms, and expanding the boundaries of what is possible.

5. Conclusions

The primary contribution of this publication is the valuable BUU-LSPINE dataset, which serves as an essential tool for researching the lumbar spine, LSTV, and spondylolisthesis. The dataset comprises records from 3600 patients, including 621 cases of spondylolisthesis, providing comprehensive data for in-depth analysis.

Spondylolisthesis is a prevalent condition affecting individuals, and its symptoms may gradually manifest over time. Early detection of the disease necessitates the development of algorithms for locating lumbar vertebrae and predicting spondylolisthesis. Our dataset offers essential information for the development and research of such algorithms.

The dataset is available at https://services.informatics.buu.ac.th/spine/ (accessed on 26 July 2023). In future studies, we intend to expand the size of our dataset and improve our feature point detection and spondylolisthesis prediction algorithms. We aim to develop a CDSS that can support physicians in diagnosing spinal conditions.