Neurocomputing 566 (2024) 126946

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

Automated detection of vertebral fractures from X-ray images: A novel

machine learning model and survey of the field✩
Li-Wei Cheng a ,1 , Hsin-Hung Chou b ,1 , Yu-Xuan Cai c ,1 , Kuo-Yuan Huang d ,2 , Chin-Chiang Hsieh e ,
Po-Lun Chu c , I-Szu Cheng f , Sun-Yuan Hsieh b,g ,2 ,∗
a
Institute of Medical Informatics, National Cheng Kung University, No. 1, University Road, Tainan, 70101, Taiwan
b Department of Computer Science and Information Engineering, National Chi Nan University, No. 1, University Rd., Puli Township, Nantou, 54561, Taiwan
c Department of Computer Science and Information Engineering, National Cheng Kung University, No. 1, University Road, Tainan, 70101, Taiwan
d
Department of Orthopedics, National Cheng Kung University Hospital, College of Medicine, National Cheng Kung University, No. 138, Sheng Li
Road, Tainan, 701, Taiwan
e
Department of Radiology, Tainan Hospital, Ministry of Health and Welfare, No. 125, Jhongshan Rd., West Central Dist., Tainan, 70043, Taiwan
f
College of Medicine, National Cheng Kung University, No. 1, University Road, Tainan, 70101, Taiwan
g
Department of Computer Science and Information Engineering, Institute of Medical Information, Institute of Manufacturing Information and Systems, National
Cheng Kung University., No. 1, University Road, Tainan, 70101, Taiwan

ARTICLE INFO ABSTRACT

Keywords: Vertebral fractures are a common problem and the most prevalent of thoracolumbar compression and burst
Thoracolumbar X-ray image fractures. However, vertebral fractures are difficult to diagnose: an experienced orthopedist or radiologist
Compression fracture is required to detect and determine the type of vertebral fracture. Thus, artificial intelligence methods for
Burst fracture
diagnosing vertebral fractures are clinically useful.
Vertebral body segmentation
On the basis of a review of 12 studies in the literature, the earliest of which was published in 2020, we
Machine learning model
Artificial intelligence in medicine
propose a machine learning model that detects and determines the type of vertebral fracture on the basis of
X-ray data. In this method, YOLOv4 and ResUNet are used to segment vertebral bodies from X-ray images. In
evaluation experiments, our method had a precision of 99%, 74%, and 94% in identifying healthy vertebrae,
compression fractures, and burst fractures, respectively.

1. Introduction Compression and burst fractures are the most common types of ver-
tebral fractures [5]. Burst fractures are common in high-energy trauma
Spine fractures can be caused by traumatic injury, especially in and are most commonly associated with falls and traffic accidents; 10%
patients with osteoporosis. Spine fractures tend to occur at the thora- of spine fractures are burst fractures [6]. The thoracolumbar spine,
columbar junction; thoracolumbar fractures are classified using various located at the transitional zone between the thoracic rib cage and
clinical classification systems. Some of these systems are based on the lumbar spine, is susceptible to injury and fracture because it undergoes
type and mechanism of fracture [1,2], and some are based on the state greater motion relative to other parts of the spine. A burst fracture is
of the anatomical structure and nerves affected by the fracture [3,4]. a thoracolumbar fracture involving the anterior and middle columns
The simplest and most widely used classification system is that based of the vertebral body. Burst fractures are usually unstable and often
on three-column theory proposed by Denis [2]. In this system, the spine accompanied by neurological symptoms and spinal instability. Thus,
a definitive diagnosis on the basis of computed tomography (CT) or
is divided into the anterior, middle, and posterior columns.
magnetic resonance imaging (MRI) is necessary [7], and immediate

✩ An extended abstract, under the title ‘‘Automated Diagnosis of Vertebral Fractures Using Radiographs and Machine Learning,’’ appeared in the Proceedings
of the 18th International Conference on Intelligent Computing, ICIC 2022, August 2022, Xi’an, China. LNCS 13393, pp.726-738, 2022.
∗ Corresponding author at: Department of Computer Science and Information Engineering, Institute of Medical Information, Institute of Manufacturing
Information and Systems, National Cheng Kung University., No. 1, University Road, Tainan, 70101, Taiwan.
E-mail addresses: [email protected] (L.-W. Cheng), [email protected] (H.-H. Chou), [email protected] (Y.-X. Cai),
[email protected] (K.-Y. Huang), [email protected] (C.-C. Hsieh), [email protected] (P.-L. Chu), [email protected]
(I.-S. Cheng), [email protected] (S.-Y. Hsieh).
1
Li-Wei Cheng, Hsin-Hung Chou, and Yu-Xuan Cai contributed equally to this work.
2
Kuo-Yuan Huang and Sun-Yuan Hsieh contributed equally to this work.

https://doi.org/10.1016/j.neucom.2023.126946
Received 25 April 2023; Received in revised form 4 September 2023; Accepted 16 October 2023
Available online 30 October 2023
0925-2312/© 2023 Elsevier B.V. All rights reserved.
L.-W. Cheng et al. Neurocomputing 566 (2024) 126946

surgery may even be required. However, a resident or emergency room

physician may not be available for timely diagnosis and may have
difficulty identifying thoracolumbar fractures or even differentiating
between burst and compression fractures on the basis of X-ray im-
ages. These impediments to timely and accurate diagnosis severely
worsen the affected patient’s prognosis. Thus, artificial intelligence (AI)
methods for diagnosing burst fractures are required.
The use of deep learning for vertebral fracture detection remains
nascent, and existing methods involve processing X-ray, CT, or MRI
data to obtain a diagnosis [8–14]. Generally, morphological patterns
can be more clearly represented in CT or MRI images than in X-
ray images. However, unlike X-ray screenings, CT and MRI screenings
are more expensive, more time-consuming, and thus less clinically
accessible. Furthermore, patients are subject to high levels of radiation Fig. 1. Number of publications each year on AI-assisted vertebral fracture detection.

if CT is used to extensively screen the spine for fracture.

Thus, machine learning methods should be capable of processing
X-ray data, CT data, and MRI data to take advantage of the benefits pixels are considered and the image is divided into the target, back-
(i.e., accuracy vs. accessibility) of each. However, existing methods can ground, and edge for segmentation. The level-set method has also been
process only X-ray data or only either CT or MRI data. Only one study applied, wherein pixels on a two-dimensional plane are mapped to
thus far trained their model on X-ray data and tested it on MRI and CT points in three-dimensional space to determine the shape and boundary
data [15]. Furthermore, machine learning methods for differentiating of an object. CNNs represent the state of the art in image segmentation,
between burst and compression fractures in the thoracolumbar region with U-Net being the most commonly used in medical applications. The
on the basis of X-ray images are lacking. To address these problems, model used in this study is a modification of U-Net called ResUNet [32–
we constructed a machine learning model that can detect compression 34].
fractures and burst fractures and differentiate between the two using Vertebral fractures are typically detected by machine learning mod-
X-ray images – a first in the literature – and we tested our model on CT els on the basis of medical images obtained using CT, MRI, or X-ray.
and MRI data. In contrast to previous studies [8–12,14], we combined MRI images display pathological structures and can clearly highlight
several models and accounted for various problem scenarios to improve the structural relationships between bone, muscle tissue, and water.
the applicability of our model. Akito et al. [12] formulated a model that identifies new osteoporotic
In addition to presenting our novel method, we present a survey vertebral fractures in MRI images. Their data set of 1624 MRI slices
on the state of the art of deep learning methods for vertebral fracture from 814 patients was annotated by two spine surgeons; among these
diagnosis on the basis of X-ray data. This survey covered 12 works, the 1624 slices, 648 and 976 slices had the position of the vertebral body
earliest of which was published in 2020. marked by a human expert and by the model, respectively. Their binary
classification model was then trained to identify old and new fractures
2. Literature review for each vertebral body, and it achieved a accuracy rate of 88%.
However, MRI imaging is slow and expensive and has poorer per-
Many early feedforward neural networks had only one hidden layer. formance than CT in imaging related to bone diseases; thus, CT is pre-
In a previous study, the extreme learning machine (ELM) algorithm was ferred. The 𝑘-nearest-neighbor method of Hamid et al. [8] preprocesses
used to increase learning efficiency in these early networks [16]. In the CT data, segments the lumbar vertebra, determines the boundaries of
ELM algorithm, a randomly selected input weight is biased to obtain an the vertebral body, and extracts vertebral body features. In preprocess-
output weight. The radial basis function (RBF) was also used in these ing, the slice image is denoised using a two-dimensional Gaussian filter,
early networks [17–19]. The neurons of the hidden layer apply the RBF and this is done to select the best slice in a CT image. The lumbar spine
on input data, and the output is passed to the output layer. These early is segmented using the open-source Insight Toolkit. The boundary of the
networks are suitable for handling large data sets, and improvements vertebral body is located at the upper and lower points of the anterior,
have been made in terms of their architecture and parameters [20–25]. center, and posterior parts of the vertebral body for a total of six points.
However, neural networks have become deeper with improvements In evaluation experiments, the correct rate was 87.5%.
in technology; convolutional neural networks (CNNs) are examples of Fig. 1 presents the number of publications each year on AI-assisted
deep networks. CNNs have been used for image segmentation, but vertebral fracture detection.
image identification is also necessary in practice. Thus, researchers Unlike CT and MRI, X-ray is quick and inexpensive; however, X-
have developed deep networks for object detection [26–30]. In early ray images do not clearly distinguish between tissues, muscles, and
object detection methods, sliding window detection is applied, where fractures, and bones may overlap on images. Several machine learn-
image detection is applied to subregions (i.e., windows) of the image. ing methods that diagnose fractures using X-ray images have been
However, this method has a long computational time because objects proposed. Tables 1 and 2 summarize studies from 2020 to 2022 and
differ in size. Similarly, in region-based CNN (R-CNN), object detection detail the year of publication, weakness, the task undertaken, the data
and feature extraction are applied to subregions of an image. Although set used for training and evaluation, the size of the data set, the
R-CNNs have much better detection speeds compared with earlier models used, and the performance results from evaluation experiments.
methods, they are not suitable for some applications, such as objection In general, most related studies have focused on detecting general
detection from video data. Thus, we chose YOLO (you only look once) vertebral fractures, osteoporotic fractures, or compression fractures.
for object detection in our CNN because it strikes the best balance Murata et al. [14] formulated a deep CNN and collected plain
between accuracy and detection speed [31]. As its name suggests, radiography data of the thoracolumbar region from 300 patients be-
YOLO’s recognition speed is fast because only one forward pass is tween 2015 and 2018; 50% and 50% of the data set comprised images
required to produce a final output. with and without vertebral fractures, respectively. The model had a
Our method also segments parts of the vertebral section. Early posttraining correct rate of 86% and F1 score of 86%.
segmentation methods have involved threshold segmentation, region Kim et al. [13] proposed a structured method for detecting com-
segmentation, or segmentation based on texture and edge features— pression fractures of the lumbar spine. The method proceeds as fol-
wherein the gray value distribution of a pixels and its neighboring lows. First, X-ray images are subject to preprocessing, where adaptive

2
L.-W. Cheng et al. Neurocomputing 566 (2024) 126946

Table 1
Recent studies on the detection of vertebral fractures using artificial intelligence models.
Reference (year) Task Dataset Size of Models used Performance Weaknessa
dataset
Murata et al. Detection of vertebral a level-3 trauma 300 DCNN Accuracy, sensitivity, specificity = (A)
(2020) [14] fractures medical center 86.0%, 84.7%, 87.3%
Kim et al. Vertebra detection and Severance Hospital, 797 M-net Mean Dice similarity metric of (A)(B)(C)
(2021) [13] segmentation Yonsei University 91.60 ± 2.22%
College of Medicine
Li et al. (2021) Detection of vertebral Taipei Veterans General 941 You Only Look Once AI model with ensemble method (A)(D)
[35] fractures Hospital version 3 (YOLOv3) => Accuracy: 93%,
sensitivity:91%, specificity: 93%;
AI model on bootstrapping
method were 89%, 83%, 95%.
Kim et al. Measurement the Gachon Gil Hospital 339 Multi-dilated recurrent Sensitivity: 0.937, specificity: (A)(B)(C)
(2021) [36] vertebral compression residual U-Net 0.995, accuracy: 0.992, dice
ratio (MDR2-UNet) similarity coefficient: 0.929, AUC:
0.987, precision–recall curve:
0.916
Chen et al. Identification of National Taiwan 1306 DCNN Accuracy: 73.59%, sensitivity: (A)
(2021) [37] vertebral fractures University Hospital 73.81%, specificity: 73.02%, AUC:
HsinChu Branch 0.72
a (A) The classification task remains simplistic, limited to identifying the presence or absence of fractures, or distinguishing a singular fracture type, thus precluding the recognition
of various fracture types; (B) The detection of vertebrae solely pertains to the lumbar region, omitting considerations for the thoracic vertebrae segment. This exclusion overlooks
the susceptibility of older individuals to experiencing fractures in the thoracic region or harboring the risk of existing fractures; (C) Within the computation of vertebral height
compression ratios, the oversight of consecutive vertebral fractures, or fractures stemming from compression exclusively in the mid or posterior segments of the vertebrae.
Consequently, the computed values of height compression ratios are imprecise, impeding the detection of fractures; (D) During height measurements or when utilizing the Genant
method for fracture detection, the omission of scenarios where the anterior, middle, and posterior sections of the vertebrae experience proportionate compression. In such instances,
the outcomes of anterior, middle, and posterior height ratio calculations fail to detect fractures.

contrast enhancement is applied using adoptive histogram equaliza- and older adult patients and recommended that different models be
tion. Second, pose-driven learning is used to identify each of the five used for young versus older adult patients. Their method achieved a
lumbar vertebrae. Third, a multistream network (M-net) is used to final accuracy of 93.36%.
segment the individual lumbar vertebra. Finally, the segmentation re- Experts have recently set a range for the determination of vertebral
sults are fine-tuned by being combined using the level-set method. The fractures, with osteoporotic compression fractures being the most com-
PoseNet, M-net, and level-set methods are combined for segmentation. mon. Fractures in older adult patients are usually osteoporotic fractures
In evaluation experiments, this combined method far outperformed or osteoporosis-induced compression fractures; data on these fractures
the traditional U-net or M-net methods, achieving a Dice coefficient, are frequently used in clinical tests, but the clinical detection rate is
precision, and specificity of 91.6±2.22%, 84.57±3.64%, and 99.5±0.17%, low. Hong et al. [41] tested a model on clinical data and on 29,307
respectively. X-ray images of the lateral spine; this data set was collected between
Kim et al. [36] proposed a multidilated recurrent residual U-Net 2007 and 2018 and covered 10,341 patients aged 40 years or older.
(MDR2-UNet) architecture for vertebral segmentation; the architecture Their model had F1 scores of 0.92 and 0.78.
features a multidilated residual block and recurrent residual block, and Dong et al. [43] formulated a GoogLeNet model that detects (1)
dilated convolution is used to increase the receptive field. In that study, moderate to severe fractures and (2) normal, trace, or mild fractures—
339 X-ray images were divided into training, validation, and testing
defined on the basis of the Genant standard. They used a data set
sets at a ratio of 6:2:2. In evaluation experiments, the aforementioned
comprising spine radiographs collected from 19,985 male patients, and
method had a final Dice similarity coefficient of 92.9%.
100,409 vertebral bodies were extracted. In evaluation experiments,
Chen et al. [37] proposed a deep CNN for fracture detection. They
the model achieved a sensitivity of 59.8% and an F1 score of 0.72.
collected data from 1306 patients from National Taiwan University
Chen et al. [42] trained and evaluated an ensemble model on ground
Hospital (NTUH), Hsinchu Branch, from 2015 to 2018, and the data
truth MRI data that were of newly formed compression fractures or
were labeled using the semiquantitative method of Genant. ImageNet
old fractures. In evaluation experiments, the performance results of
was used for pretraining, and Grad-CAM was used to visualize the
the model for the identification of newly formed vertebral compression
attention heatmap of the model for transparency. Different from other
fractures were AUC = 0.80, accuracy = 74%, sensitivity = 80%, and
studies, this study used plain radiographic data of the frontal abdominal
specificity = 68%.
region for training.
Li et al. (2021) [35] tested YOLO version 3 (YOLOv3) against human Kong et al. [39] proposed a CNN-based DeepSurv algorithm that
experts in vertebral fracture detection on a data set of 941 plain lateral predicts the occurrence of a fracture on the basis of lumbar spine X-
radiographs from 941 older adult patients (average age: 76 ± 12 years); ray images and clinical characteristics, such as age, gender, weight,
this data set was collected between 2016 and 2018 and featured 1101 glucocorticoid use, and secondary osteoporosis status. The training
vertebral fractures in the thoracic vertebrae T7 to lumbar vertebrae L5. data covered mostly middle-aged to older adult women (74.4% female;
YOLOv3 had a final accuracy of 93%. average age: 60.5 years). The model outperformed the Fracture Risk
In the method of Chou et al. (2022) [15], YOLOv3 is used for Assessment Tool (FRAX) and a Cox proportional hazards model.
vertebral body localization and segmentation. The contrast and size of Rosenberg et al. [40] focused on the detection of traumatic thora-
the images are adjusted as part of data preprocessing, and an ensemble columbar fractures, which is difficult because the presence of many
model is used to determine the presence and grade of fractures per the organs in this chest region complicates fracture detection. ResNet18
Genant classification scheme; the fracture is graded on the basis of in- and VGG16 was used for fracture prediction and comparison, respec-
jury height at the anterior, middle, and posterior thirds of the vertebral tively. The aforementioned research group used a data set comprising
body into grade 1 (<25%), grade 2 (26% to 40%), and grade 3 (>40%) 630 sagittal radiographs of the vertebra of 151 patients; vertebral frac-
fractures. Chou et al. tested their model on data from young patients tures were present and absent in 302 and 328 radiographs, respectively.

3
L.-W. Cheng et al. Neurocomputing 566 (2024) 126946

Table 2
Recent studies (published in 2022) on the detection of vertebral fractures using artificial intelligence models.
Reference (year) Task Dataset Size of Models used Performance Weakness
Dataset
Chou et al. Vertebral fracture Private dataset from 1 941 YOLOv3, ResNet34, older adult population: accuracy, (A)(D)
(2022) [15] detection centers DenseNet121, sensitivity, specificity = 93.36%,
DenseNet201 88.97%, 94.26% younger adult
population: 93.75%, 65.00%,
98.49%
Xiao et al. Detection of Private dataset from 15 5970 Ofeye 1.0 Specificity of 97.1%, a sensitivity (A)(D)
(2022) [38] compressive vertebral centers of 86%, and an accuracy of
fracture 93.9%
Kong et al. Predicting osteoporotic Seoul National 1595 DeepSurv (CNN-based) C-index values => DeepSurv, (A)
(2022) [39] fracture University Hospital 0.612; FRAX, 0.547; CoxPH,
0.594
Rosenberg et al. Detection of traumatic Spine Surgery 630 ResNet18 and VGG16 ResNet18: sensitivity (91%), (A)
(2022) [40] thoracolumbar fractures Reference Center specificity (89%), and accuracy
(88%); VGG16 = 90%, 83%, 86%
Hong et al. Detection of vertebral Severance Hospital, 29 307 DNN VF => AUROC = 0.94, sensitivity (A)
(2022) [41] fractures and Seoul 0.88, specificity 0.85; F1 score
osteoporosis 0.92Osteoporosis => AUROC =
0.86, sensitivity 0.81, specificity
0.73; F1 score 0.78
Chen et al. Identifying vertebral The Second Affiliated 1099 DL model AUC: 0.80; accuracy: 74%; (A)
(2022) [42] compression fractures Hospital of Chongqing sensitivity: 80%; specificity: 68%
Medical University
Dong et al. Detection of The Osteoporotic 19 985 GoogLeNet sensitivity of 59.8%, a PPV of (A)(D)
(2022) [43] osteoporotic Fractures in Men 91.2%, F score of 0.72.AUC-ROC
compression fractures and the precision–recall curve
were 0.99, and 0.82

ResNet18 had the highest performance in evaluation experiments and In YOLO [46], a neural network processes an entire image once to
achieved 91% sensitivity and 88% accuracy. determine the location and confidence of bounding boxes and the cate-
In addition to being difficult to characterize, osteoporotic or com- gory to which the object in the bounding box belongs. The architecture
pression fractures in the lumbar spine and thoracic spine are charac- of the YOLOv4 model is presented in Fig. 3. The confidence score of the
terized in a similar manner. Thus, Xiao et al. [38] proposed a system model is provided in Fig. 5.
that detects and classifies osteoporosis-related compressive vertebral U-net [48] is useful for biomedical image segmentation [49–51],
fractures on lateral chest radiographs; these fractures are difficult to and Resnet [52] has achieved outstanding performance in image recog-
detect in practice. They applied their system to data on older adult nition. ResUNet [44], in which residual blocks are combined with a U-
women. Fractures with a vertebral height loss of <25%, 25%–40%, net architecture (Fig. 4), is a combination of both methods. The residual
40%–67%, and >67% were characterized as mild, moderate, severe, blocks, also known as ResBlocks, can prevent gradient diffusion.
and vertebral collapse fractures, respectively. Their system had 93.9% Generally, the size of the receptive field is closely related to the
accuracy in evaluation experiments. amount of contextual information that a CNN can capture. Because
the amount of global information is usually insufficient, global aver-
3. Methods age pooling is used to provide more global information to the CNN.
However, in some data sets, spatial relationships may be lost, causing
We present our method for detecting fractures and distinguishing ambiguity. Thus, Zhao et al. [53] proposed a method termed hierar-
between compression and burst fractures on the basis of X-ray images. chical global priority, name is Pyramid Scene Parseing Network. This
We applied our method to lateral X-ray images on the thoracolum- method features from different levels generated by pyramid pooling
bar region. In this method, X-ray images are preprocessed, YOLO are smoothly connected into a fully connected layer for classification,
version 4 (YOLOv4) [31] is used for preliminary segmentation, and allows for CNN classification for different sized images and reduces the
ResUNet [44] is used for actual segmentation. The random forest ap- loss of information on the relationship between different regions.
proach is then used to determine whether the image shows no fracture, After hierarchical global pooling is applied, vertebral bodies with
a burst fracture, or a compression fracture. a screw or with bone cement can be detected (Fig. 6). Parts of the
vertebral with a screw and with no prior surgery are marked in cyan
3.1. Model architecture and blue, respectively.

To solve the problem of low contrast in X-ray images, we used adap- 3.2. Feature analysis
tive histogram equalization in data preprocessing [13,45]. To solve the
problem of input images having different sizes, we used YOLOv4 [31] The lateral X-ray images of the thoracolumbar region in the data set
for preliminary segmentation. The overall model architecture is pre- were annotated by an orthopedic doctor with regard to whether a burst
sented in Fig. 2. or compression fracture was present. These fractures can be diagnosed
Adaptive histogram equalization [45] is a method for contrast en- on the basis of the height and proportion of the anterior, middle, and
hancement with broad applicability and demonstrable effectiveness. A posterior columns of the vertebral body. The ratios of the corresponding
previous study demonstrated that a machine learning network can iden- heights of the lower or upper vertebral bodies must also be consid-
tify the location of each thoracolumbar vertebral body in input images ered. We extracted features from images that we segmented using our
of different sizes [13]. Therefore, we used the YOLOv4 model [31], method, and these features were used to train machine learning models
which is fast and accurate, to preliminary segment vertebral bodies that identify the vertebral body type. We evaluated these models and
from T1 to L5 in the thoracolumbar region. used the best performing model as our feature analysis model.

4
L.-W. Cheng et al. Neurocomputing 566 (2024) 126946

Fig. 2. Workflow of the proposed method.

Fig. 3. YOLO model architecture [47].

Compression fractures and burst fractures are primarily distin- If a vertebral body collapses, its anterior height is expected to
guished by the presence of damage to the middle column of the be discontinuous with the vertebral body above or below it. In this
vertebral body; they are also distinguished by the height of the an- scenario, a compressed and burst vertebral fracture can be detected. If
terior and posterior vertebral bodies. The vertebral bodies are the the posterior height of a vertebral body is similar to that of the vertebral
main weight-bearing structures in the human body, and their heights body above or below it, the middle column may be undamaged; in this
typically increase gradually from top to bottom as a result of healthy situation, a burst fracture diagnosis can be excluded. If the ratio of the
growth. Therefore, physicians should first measure the anterior and anterior and posterior height of this vertebral body is less than 0.8 and
posterior heights of three consecutive vertebral bodies, including the the posterior height has not collapsed, a compression fracture is likely
fractured vertebral body, when determining whether the fracture is a to have occurred. Normally, the height of the posterior vertebral body
compression fracture or a burst fracture. gradually increases from top to bottom. If it does not gradually increase

5
L.-W. Cheng et al. Neurocomputing 566 (2024) 126946

Fig. 4. ResUNet model architecture.

Fig. 6. Identification of vertebral bodies with screw by the model.

the model calculates the ratio of 𝐻𝑚 to 𝐻𝑎, denoted as 𝑅_𝐻𝑚𝐻𝑎; the

ratio of 𝐻𝑝 to 𝐻𝑎, denoted as 𝑅_𝐻𝑝𝐻𝑎; and the ratio of 𝐻𝑚 to 𝐻𝑝,
denoted as 𝑅_𝐻𝑚𝐻𝑝. These features are illustrated in Table 3.

3.3. Data

We used data from 390 patients receiving orthopedic care (302

Fig. 5. Confidence scores output by the model. female and 88 male patients) at our hospital; we obtained human
study approval from the Institutional Review Board of our hospital.
The data set comprised 390 thoracolumbar X-ray images obtained
or is even lower than the posterior height of the upper vertebral from a hospital-wide database; these pictures were taken at any time
body, then the middle column of the vertebral body is damaged and between January 2014 and December 2020. The patients were aged
collapsed, and a burst fracture is likely to have occurred. 27 to 91 years, and the mean age was 75.12 ± 10.03 years. We were
We used features that are generally used to diagnose vertebral frac- aware that the presence of single-level or multilevel spinal fractures
tures; these features in segmented images are calculated by the model could compromise the integrity of the experiment. In total, 271, 93,
as follows. First, the model measures the anterior (𝐻𝑎), middle (𝐻𝑚), and 26 patients had fractures at 1, 2–4, and ≥5 levels, respectively. The
and posterior (𝐻𝑝) heights of vertebral bodies (Fig. 7). Subsequently, demographic characteristics of the patients are presented in Table 4.

6
L.-W. Cheng et al. Neurocomputing 566 (2024) 126946

Table 4
Demographic characteristics of patients.
Patients with vertebral fracture 390
Age (years old) 75.12 ± 10.03
Gender: no. patients (%)
Male 88 (22.6)
Female 302 (77.4)
Single/Multi-level spinal fractures: no. patients (%)
Single level 271 (69.5)
2-4 level 93 (23.8)
5+ level 26 (6.7)

Fig. 7. Anterior, middle, and posterior heights of a vertebral body.

Table 3
Features of segmented images.
Attribute Description
𝐻𝑎 Anterior height of the vertebral body
𝐻𝑚 Middle height of the vertebral body
𝐻𝑝 Posterior height of the vertebral body
𝐻𝑚
𝑅_𝐻𝑚𝐻𝑎 Ratio of 𝐻𝑚 and 𝐻𝑎 (= ) Fig. 8. Vertebral body labels.
𝐻𝑎
𝐻𝑝
𝑅_𝐻𝑝𝐻𝑎 Ratio of 𝐻𝑝 and 𝐻𝑎 (= )
𝐻𝑎
𝐻𝑚
𝑅_𝐻𝑚𝐻𝑝 Ratio of 𝐻𝑚 and 𝐻𝑝 (=
𝐻𝑝
) executed over five iterations. At each iteration, a different subset is
𝐻𝑖
used for testing. The mean performance result is used to indicate the
𝑅_𝐻𝑖_𝑙𝑜𝑤𝑒𝑟 = , performance of the model.
𝐻𝑖 𝑜𝑓 𝑡ℎ𝑒 𝑎𝑑𝑗𝑎𝑐𝑒𝑛𝑡 𝑙𝑜𝑤𝑒𝑟 𝑣𝑒𝑟𝑡𝑒𝑏𝑟𝑎𝑙 𝑏𝑜𝑑𝑦
𝐻𝑖 = 𝐻𝑎, 𝐻𝑚, 𝑜𝑟 𝐻𝑝 We use Python [56] as the programming language. Experiments
(=−1 when no lower vertebral body is present)
were conducted in TensorFlow [57], and Keras [58], and image pro-
𝐻𝑖 cessing was implemented using OpenCV [59] and VGG Image Annota-
𝑅_𝐻𝑖_𝑢𝑝𝑝𝑒𝑟 = ,
𝐻𝑖 𝑜𝑓 𝑡ℎ𝑒 𝑎𝑑𝑗𝑎𝑐𝑒𝑛𝑡 𝑢𝑝𝑝𝑒𝑟 𝑣𝑒𝑟𝑡𝑒𝑏𝑟𝑎𝑙 𝑏𝑜𝑑𝑦
𝐻𝑖 = 𝐻𝑎, 𝐻𝑚, 𝑜𝑟 𝐻𝑝 tor [54]. The experiments were run on a computer with an AMD Ryzen
(=−1 when no upper vertebral body is present) 5 3600 CPU, 32 GB of DDR4 RAM, and a NVIDIA GeForce RTX 2070
𝑅_𝐻𝑖_𝑙𝑜𝑤𝑒𝑟_𝑒𝑛𝑐𝑜𝑑𝑒 =1 when 𝑅_𝐻𝑖_𝑙𝑜𝑤𝑒𝑟 < 0.8 SUPER GPU.
𝑅_𝐻𝑖_𝑢𝑝𝑝𝑒𝑟_𝑒𝑛𝑐𝑜𝑑𝑒 =1 when 𝑅_𝐻𝑖_𝑢𝑝𝑝𝑒𝑟 < 0.8 The following metrics were used. First, the Dice coefficient was used
to evaluate vertebral body segmentation quality. The mean of the Dice
coefficients for all vertebral bodies was used. The Dice coefficient is
defined as follows.
Vertebral bodies at T1 to L5 were given one of four labels: normal,
2 × |𝑃 ∩ 𝐺|
compression fracture, burst fracture, and others; the ‘‘others’’ label was 𝐷𝑖𝑐𝑒 =
|𝑃 | + |𝐺|
given if the vertebral body had bone cement or a screw. The labeled
images were verified against CT or MRI images by an experienced where P denotes the predicted segmentation results, G denotes the
orthopedic doctor. VGG Image Annotator [54] software was used for corresponding ground truth segmentation, and the operator|.| returns
labeling. Examples of vertebral bodies corresponding to each of the four the number of labeled voxels.
labels are presented in Fig. 8. Vertebral bodies were also labeled on the Accuracy, precision, recall, and F1 score were used to evaluate
basis of whether they were at the upper thoracic, thoracolumbar, or diagnostic performance. These metrics are defined as follows.
lower lumbar regions [12]. The data set had images of 3634 vertebral 𝑇𝑃
𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
bodies, and the numbers of vertebral bodies corresponding to each of 𝑇𝑃 + 𝐹𝑃
the aforementioned labels are presented in Table 5.
𝑇𝑃
𝑅𝑒𝑐𝑎𝑙𝑙 =
𝑇𝑃 + 𝐹𝑁
3.4. Setup of evaluation experiments
2 × 𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 × 𝑅𝑒𝑐𝑎𝑙𝑙
𝐹 1-𝑠𝑐𝑜𝑟𝑒 =
𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙
Standard five-fold cross-validation was used in evaluation exper-
iments [55]; this method allows for all observations to be used for 𝑇𝑃 + 𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
training and testing and for each observation to be used for testing 𝑇𝑃 + 𝐹𝑃 + 𝑇𝑁 + 𝐹𝑁
only once. In this method, the data set is split into five subsets – where TP, FP, TN, and FN represent the numbers of true positives, false
four for training, and one for testing – and training and testing are positives, true negatives, and false negatives, respectively.

7
L.-W. Cheng et al. Neurocomputing 566 (2024) 126946

Table 5
Number of vertebral bodies for each label.
Total Labeled with:
Normal Compression Burst Others
Upper thoracic level (T1-T9) 729 634 38 41 16
Thoracolumbar level (T10-L2) 1809 1268 171 288 82
Lower lumbar level (L3-L5) 1096 979 16 3 98
Total 3634 2881 225 332 196

Fig. 9. CT or MRI images used for verification.

We used a multiclass confusion matrix to present the diagnostic We compared our model against support vector machine, extreme
results. The model outputs from X-ray images were verified against gradient boosting (XGBoost), random forest, multilayer perceptron, and
model outputs from CT or MRI images; CT and MRI images of a 𝑘-nearest-neighbor models. Their performance results are presented in
compression fracture and burst fracture are presented in Fig. 9. This Table 6 and Fig. 11 for comparison. We used a multiclass confusion
was done to verify the feasibility of applying this model to X-ray images matrix to present our results (Fig. 12). We chose the random forest
for diagnosis. model for feature analysis after the models were compared. Our model
had precision scores of 99%, 74%, and 94% for the identification of
4. Results no fractures, compression fractures, and burst fractures, respectively.
For the determination of whether a fracture (in general) was present
We used 390 vertebral X-ray images in the sagittal view of 3634 or absent, our model had an accuracy, precision, recall, and F1-score
vertebral bodies in evaluation experiments. Manual and model segmen- of 92.0%, 93.2%, 95.7% and 94.4%, respectively.
tation results are presented in Fig. 10 for comparison. The average Dice Our model took 30 s to output an annotated image, taking 1–2 s for
coefficient for segmentation was 0.852. each step.

8
L.-W. Cheng et al. Neurocomputing 566 (2024) 126946

Fig. 10. The images from left to right are original, manually segmented, and model segmented images, respectively. The rightmost image has the manual segmentation and model
segmentation results overlaid on top of each other.

Table 6
Performance results of models for burst or compression fracture identification in the thoracolumbar region.
Total Normal
Model Accuracy Precision Recall F1-score
RF 0.98 0.94 0.99 0.96
KNN 0.89 0.90 0.98 0.94
SVM 0.89 0.91 0.98 0.94
MLP 0.89 0.89 0.99 0.94
XGBoost 0.96 0.98 0.99 0.99
Compression Burst
Model Precision Recall F1-score Precision Recall F1-score
RF 0.77 0.74 0.75 0.86 0.94 0.90
KNN 0.75 0.71 0.73 0.93 0.87 0.90
SVM 0.60 0.50 0.55 0.86 0.70 0.78
MLP 0.83 0.45 0.59 0.91 0.69 0.78
XGBoost 0.85 0.73 0.79 0.86 0.86 0.86

Fig. 11. Performance results of various models in diagnosing burst or compression fractures in the thoracolumbar region.

5. Discussion stability, reducing nerve compression, preventing neurological deterio-

ration, realigning the spine to prevent deformity, and promoting early
The thoracolumbar spine is located in the transition zone be- pain relief [61].
tween the thoracic and lumbar vertebrae; thoracolumbar fractures are Our data set covered relatively old patients. Older patients are more
common because this region is subject to relatively greater motion than likely to have fractures following low-energy trauma (i.e., trauma due
other regions. Thoracolumbar fractures are classified as burst fractures to a fall from a standing height), which may cause spinal deformity,
if they involve the anterior and medial columns of the vertebral body. because their bones tend to be more brittle due to osteoporosis. Thus,
Surgery should be considered in cases of unstable burst fractures that because fractures following low-energy trauma are less difficult to
are accompanied by (1) posterior ligamentous complex injury, (2) a detect than other fractures, our model is likely to perform well in
comminuted vertebral body with a vertebral height collapse of more diagnosing more obvious major-trauma fractures, such as those caused
than 50%, (3) neurological deficits, (3) lumbar spinal canal compromis- by traffic accidents or a fall from a height.
ing more than 50%, (4) thoracic spinal canal compromise of more than Our results indicated that scoliosis or multilevel fractures hindered
30% as a result of a retropulsed bony fragment, and (5) deformities the model’s ability to distinguish between fractures. In cases of contin-
such as (5a) kyphosis (>30◦ ) or (5b) scoliosis (>20◦ ) [60]. Thus, the uous, multilevel fractures, the height of vertebral bodies adjacent to a
objectives of surgery for thoracolumbar burst fractures are restoring fractured body – a key indicator used by our model – is not indicative of

9
L.-W. Cheng et al. Neurocomputing 566 (2024) 126946

Nonetheless, our results demonstrate the clinical utility of our

model. Furthermore, as indicated in our survey, the quality and quan-
tity of training data are key to model performance. Thus, multivariate
data or large data sets can be used to improve our model.

Declaration of competing interest

The authors declare the following financial interests/personal rela-

tionships which may be considered as potential competing interests:
Sun-Yuan Hsieh reports was provided by National Cheng Kung Univer-
sity. Sun-Yuan Hsieh reports a relationship with National Cheng Kung
University that includes: employment.

Data availability

Data will be made available on request.

Fig. 12. Confusion matrix of the random forest model.
Acknowledgment
Table 7
Performance results of our method and those of other studies for comparison. This research was partially supported by National Science and
Accuracy Precision Recall F1-score Technology Council, grant number of MOST 111-2221-E-260-008-MY2
Kazuma Murata et al. 0.860 0.873 0.847 0.860
(Hsin-Hung Chou).
-for vertebral fracture (X-ray)
This study 0.920 0.932 0.957 0.944 References
-for vertebral fracture (X-ray)
[1] F. Magerl, M. Aebi, S. Gertzbein, J. Harms, S. Nazarian, A comprehensive
classification of thoracic and lumbar injuries, Eur. Spine J. 3 (4) (1994) 184–201.
[2] F. Denis, The three column spine and its significance in the classification of acute
fracture. In cases of scoliosis, our model may perform poorly if skewed thoracolumbar spinal injuries, Spine 8 (8) (1983) 817–831.
vertebral bodies are present in the sagittal view of X-ray images, and [3] K.K. Haussler, Anatomy of the thoracolumbar vertebral region, Vet. Clin. North
Amer.: Equine Pract. 15 (1) (1999) 13–26.
CT or MRI may be required instead. These limitations can be addressed
[4] G.J. Groen, B. Baljet, J. Drukker, Nerves and nerve plexuses of the human
in future studies.
vertebral column, Amer. J. Anat. 188 (3) (1990) 282–296.
Our method is similar to that of Kim et al. [13], which is used in [5] J.W. Savage, G.D. Schroeder, P.A. Anderson, Vertebroplasty and kyphoplasty for
combination with the level-set method. Specifically, their model uses the treatment of osteoporotic vertebral compression fractures, JAAOS J. Amer.
pose-driven learning to identify the five lumbar vertebrae and M-net Acad. Orthop. Surg. 22 (10) (2014) 653–664.
to segment individual vertebrae. In evaluation experiments involving [6] F.V. Bensch, M.P. Koivikko, M.J. Kiuru, S.K. Koskinen, The incidence and
clinical data, their model had a center-position detection error rate of distribution of burst fractures, Emerg. Radiol. 12 (3) (2006) 124–129.
25.35 ± 10.86% and Dice similarity coefficient of 91.60 ± 2.22%. Our [7] L.-Y. Dai, S.-D. Jiang, X.-Y. Wang, L.-S. Jiang, A review of the management of
thoracolumbar burst fractures, Surg. Neurol. 67 (3) (2007) 221–231.
method is also similar to the deep CNN of Murata et al. [14] for
[8] H. Yousefi, E. Salehi, O.S. Sheyjani, H. Ghanaatti, Lumbar spine vertebral
detecting vertebral fractures from plain radiographs of the spine. Their
compression fracture case diagnosis using machine learning methods on CT
method had an accuracy, sensitivity, and specificity of 86.0%, 84.7%, images, in: 2019 4th International Conference on Pattern Recognition and Image
and 87.3%, respectively, in evaluation experiments. Analysis (IPRIA), IEEE, 2019, pp. 179–184.
We are the first to propose a method that can distinguish be- [9] M. Pisov, V. Kondratenko, A. Zakharov, A. Petraikin, V. Gombolevskiy, S.
tween compression and burst fractures. The performance results of our Morozov, M. Belyaev, Keypoints localization for joint vertebra detection and
model and its state-of-the-art counterparts are presented in Table 7 for fracture severity quantification, in: International Conference on Medical Image
comparison. Computing and Computer-Assisted Intervention, Springer, 2020, pp. 723–732.
[10] J. Nicolaes, S. Raeymaeckers, D. Robben, G. Wilms, D. Vandermeulen, C.
Libanati, M. Debois, Detection of vertebral fractures in CT using 3D convolutional
6. Conclusions
neural networks, in: International Workshop and Challenge on Computational
Methods and Clinical Applications for Spine Imaging, Springer, 2019, pp. 3–14.
In this study, we surveyed AI methods that detect vertebral fractures [11] T. Li, B. Wei, J. Cong, X. Li, S. Li, S 3 egANet: 3D spinal structures segmentation
on the basis of X-ray images. We also proposed our method that is via adversarial nets, IEEE Access 8 (2019) 1892–1901.
capable of distinguishing between burst or compression fractures in [12] A. Yabu, M. Hoshino, H. Tabuchi, S. Takahashi, H. Masumoto, M. Akada, S.
the thoracolumbar region using X-ray images. In our model, YOLOv4 Morita, T. Maeno, M. Iwamae, H. Inose, et al., Using artificial intelligence to
and ResUNet are used to segment vertebral bodies, and features that diagnose fresh osteoporotic vertebral fractures on magnetic resonance images,
Spine J. (2021).
are extracted from segmented images are analyzed using a random
[13] K.C. Kim, H.C. Cho, T.J. Jang, J.M. Choi, J.K. Seo, Automatic detection and
forest model. Our method outperformed its state-of-the-art counterparts
segmentation of lumbar vertebrae from X-ray images for compression fracture
in evaluation experiments, achieving an average Dice coefficient of evaluation, Comput. Methods Programs Biomed. 200 (2021) 105833.
85.2% for segmentation and precision scores of 99%, 74%, and 94%, [14] K. Murata, K. Endo, T. Aihara, H. Suzuki, Y. Sawaji, Y. Matsuoka, H. Nishimura,
respectively, for identifying no fractures, compression fractures, and T. Takamatsu, T. Konishi, A. Maekawa, et al., Artificial intelligence for the
burst fractures. detection of vertebral fractures on plain spinal radiography, Sci. Rep. 10 (1)
If our model predicts a burst fracture, CT or MRI can be arranged (2020) 1–8.
to obtain a definite diagnosis and to determine whether surgery is [15] P.-H. Chou, T.H.-T. Jou, H.-T.H. Wu, Y.-C. Yao, H.-H. Lin, M.-C. Chang, S.-T.
Wang, H.H.-S. Lu, H.-H. Chen, Ground truth generalizability affects performance
necessary. Furthermore, if our model predicts a compression fracture,
of the artificial intelligence model in automated vertebral fracture detection on
first aid with conservative treatment can be applied to conserve medical plain lateral radiographs of the spine, Spine J. 22 (4) (2022) 511–523.
resources. Nonetheless, the mechanism and energy of the injury and [16] S. Li, Z.-H. You, H. Guo, X. Luo, Z.-Q. Zhao, Inverse-free extreme learning
the presence of neurological deficits or severe soft tissue damage may machine with optimal information updating, IEEE Trans. Cybern. 46 (5) (2016)
complicate the application of our model to clinical decision-making. 1229–1241.

10
L.-W. Cheng et al. Neurocomputing 566 (2024) 126946

[17] D.-S. Huang, Systematic Theory of Neural Networks for Pattern Recognition, Vol. [42] W. Chen, X. Liu, K. Li, Y. Luo, S. Bai, J. Wu, W. Chen, M. Dong, D. Guo,
201, Publishing House of Electronic Industry of China, Beijing, 1996. A deep-learning model for identifying fresh vertebral compression fractures on
[18] D.-s. Huang, Radial basis probabilistic neural networks: Model and application, digital radiography, Eur. Radiol. 32 (3) (2022) 1496–1505.
Int. J. Pattern Recognit. Artif. Intell. 13 (07) (1999) 1083–1101. [43] Q. Dong, G. Luo, N.E. Lane, L.-Y. Lui, L.M. Marshall, D.M. Kado, P. Cawthon,
[19] D.-S. Huang, J.-X. Du, A constructive hybrid structure optimization methodology J. Perry, S.K. Johnston, D. Haynor, et al., Deep learning classification of spinal
for radial basis probabilistic neural networks, IEEE Trans. Neural Netw. 19 (12) osteoporotic compression fractures on radiographs using an adaptation of the
(2008) 2099–2115. genant semiquantitative criteria, Acad. Radiol. (2022).
[20] D.-S. Huang, W. Jiang, A general CPL-AdS methodology for fixing dynamic [44] F.I. Diakogiannis, F. Waldner, P. Caccetta, C. Wu, ResUNet-a: A deep learn-
parameters in dual environments, IEEE Trans. Syst. Man Cybern. B 42 (5) (2012) ing framework for semantic segmentation of remotely sensed data, ISPRS J.
1489–1500. Photogramm. Remote Sens. 162 (2020) 94–114.
[21] D.-S. Huang, W.-B. Zhao, Determining the centers of radial basis probabilistic [45] S.M. Pizer, E.P. Amburn, J.D. Austin, R. Cromartie, A. Geselowitz, T. Greer, B. ter
neural networks by recursive orthogonal least square algorithms, Appl. Math. Haar Romeny, J.B. Zimmerman, K. Zuiderveld, Adaptive histogram equalization
Comput. 162 (1) (2005) 461–473. and its variations, Comput. Vis. Graph. Image Process. 39 (3) (1987) 355–368.
[46] J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Unified,
[22] F. Han, Q.-H. Ling, D.-S. Huang, Modified constrained learning algorithms
real-time object detection, in: Proceedings of the IEEE Conference on Computer
incorporating additional functional constraints into neural networks, Inform. Sci.
Vision and Pattern Recognition, 2016, pp. 779–788.
178 (3) (2008) 907–919.
[47] I. Pacal, D. Karaboga, A robust real-time deep learning based automatic
[23] J.-X. Du, D.-S. Huang, G.-J. Zhang, Z.-F. Wang, A novel full structure optimization
polyp detection system, Comput. Biol. Med. (ISSN: 0010-4825) 134 (2021)
algorithm for radial basis probabilistic neural networks, Neurocomputing 70
104519, http://dx.doi.org/10.1016/j.compbiomed.2021.104519, URL: https://
(1–3) (2006) 592–596.
www.sciencedirect.com/science/article/pii/S0010482521003139.
[24] J.-X. Du, D.-S. Huang, X.-F. Wang, X. Gu, Shape recognition based on neural
[48] O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for
networks trained by differential evolution algorithm, Neurocomputing 70 (4–6)
biomedical image segmentation, in: International Conference on Medical Image
(2007) 896–903.
Computing and Computer-Assisted Intervention, Springer, 2015, pp. 234–241.
[25] W.-B. Zhao, D.-S. Huang, J.-Y. Du, L.-M. Wang, Genetic optimization of radial [49] D. Shen, G. Wu, H.-I. Suk, Deep learning in medical image analysis, Annu. Rev.
basis probabilistic neural networks, Int. J. Pattern Recognit. Artif. Intell. 18 (08) Biomed. Eng. 19 (2017) 221–248.
(2004) 1473–1499. [50] G. Litjens, T. Kooi, B.E. Bejnordi, A.A.A. Setio, F. Ciompi, M. Ghafoorian, J.A.
[26] Y. Xie, W. Zhang, C. Li, S. Lin, Y. Qu, Y. Zhang, Discriminative object tracking Van Der Laak, B. Van Ginneken, C.I. Sánchez, A survey on deep learning in
via sparse representation and online dictionary learning, IEEE Trans. Cybern. 44 medical image analysis, Med. Image Anal. 42 (2017) 60–88.
(4) (2013) 539–553. [51] J. Ker, L. Wang, J. Rao, T. Lim, Deep learning applications in medical image
[27] A. Yilmaz, O. Javed, M. Shah, Object tracking: A survey, ACM Comput. Surv. analysis, IEEE Access 6 (2017) 9375–9389.
(CSUR) 38 (4) (2006) 13–es. [52] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in:
[28] Q. Wang, L. Zhang, L. Bertinetto, W. Hu, P.H. Torr, Fast online object tracking Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
and segmentation: A unifying approach, in: 2019 IEEE/CVF Conference on 2016, pp. 770–778.
Computer Vision and Pattern Recognition (CVPR), IEEE, 2019, pp. 1328–1338. [53] H. Zhao, J. Shi, X. Qi, X. Wang, J. Jia, Pyramid scene parsing network, in:
[29] X. Lu, W. Wang, C. Ma, J. Shen, L. Shao, F. Porikli, See more, know more: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
Unsupervised video object segmentation with co-attention siamese networks, 2017, pp. 2881–2890.
in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern [54] A. Dutta, A. Zisserman, The VIA annotation software for images, audio and video,
Recognition, 2019, pp. 3623–3632. in: Proceedings of the 27th ACM International Conference on Multimedia, 2019,
[30] L. Bertinetto, J. Valmadre, J.F. Henriques, A. Vedaldi, P.H. Torr, Fully- pp. 2276–2279.
convolutional siamese networks for object tracking, in: Computer Vision–ECCV [55] R. Kohavi, et al., A study of cross-validation and bootstrap for accuracy
2016 Workshops: Amsterdam, the Netherlands, October 8-10 and 15-16, 2016, estimation and model selection, in: Ijcai, Vol. 14, Montreal, Canada, 1995, pp.
Proceedings, Part II 14, Springer, 2016, pp. 850–865. 1137–1145.
[31] A. Bochkovskiy, C.-Y. Wang, H.-Y.M. Liao, Yolov4: Optimal speed and accuracy [56] G. Van Rossum, F.L. Drake, Python 3 reference manual, 2009, https://www.
of object detection, 2020, arXiv preprint arXiv:2004.10934. python.org/.
[32] X.-F. Wang, D.-S. Huang, H. Xu, An efficient local Chan–Vese model for image [57] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G.S. Corrado,
segmentation, Pattern Recognit. 43 (3) (2010) 603–618. A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving,
[33] X.-F. Wang, D.-S. Huang, A novel density-based clustering framework by using M. Isard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Levenberg, D. Mané,
level set method, IEEE Trans. Knowl. Data Eng. 21 (11) (2009) 1515–1531. R. Monga, S. Moore, D. Murray, C. Olah, M. Schuster, J. Shlens, B. Steiner,
I. Sutskever, K. Talwar, P. Tucker, V. Vanhoucke, V. Vasudevan, F. Viégas, O.
[34] Y. Zhao, D.-S. Huang, W. Jia, Completed local binary count for rotation invariant
Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu, X. Zheng, TensorFlow:
texture classification, IEEE Trans. Image Process. 21 (10) (2012) 4492–4497.
Large-scale machine learning on heterogeneous systems, 2015, URL: https://
[35] Y.-C. Li, H.-H. Chen, H.H.-S. Lu, H.-T.H. Wu, M.-C. Chang, P.-H. Chou, Can a
www.tensorflow.org/, Software available from tensorflow.org.
deep-learning model for the automated detection of vertebral fractures approach
[58] F. Chollet, et al., Keras, 2015, https://keras.io.
the performance level of human subspecialists? Clin. Orthop. Relat. Res. 479 (7)
[59] Itseez, Open source computer vision library, 2015, https://github.com/itseez/
(2021) 1598–1612.
opencv.
[36] D.H. Kim, J.G. Jeong, Y.J. Kim, K.G. Kim, J.Y. Jeon, Automated vertebral
[60] R.F. Heary, S. Kumar, Decision-making in burst fractures of the thoracolumbar
segmentation and measurement of vertebral compression ratio based on deep
and lumbar spine, Indian J. Orthop. 41 (4) (2007) 268.
learning in X-ray images, J. Digit. Imaging 34 (4) (2021) 853–861.
[61] L. Yi, B. Jingping, J. Gele, T. Wu, X. Baoleri, Operative versus non-
[37] H.-Y. Chen, B.W.-Y. Hsu, Y.-K. Yin, F.-H. Lin, T.-H. Yang, R.-S. Yang, C.- operative treatment for thoracolumbar burst fractures without neurological
K. Lee, V.S. Tseng, Application of deep learning algorithm to detect and deficit, Cochrane Database Syst. Rev. (4) (2006).
visualize vertebral fractures on plain frontal radiographs, PLoS One 16 (1) (2021)
e0245992.
[38] B.-H. Xiao, M.S. Zhu, E.-Z. Du, W.-H. Liu, J.-B. Ma, H. Huang, J.-S. Gong, D.
Diacinti, K. Zhang, B. Gao, et al., A software program for automated compressive
vertebral fracture detection on elderly women’s lateral chest radiograph: Ofeye Li-Wei Cheng received the MS degree from the Institute
1.0, Quant. Imaging Med. Surg. 12 (8) (2022) 4259. of Medical Informatics, National Cheng Kung University
[39] S.H. Kong, J.-W. Lee, B.U. Bae, J.K. Sung, K.H. Jung, J.H. Kim, C.S. Shin, in 2021. His research interests include machine learning,
Development of a spine X-ray-based fracture prediction model using a deep algorithm design and bioinformatics.
learning algorithm, Endocrinol. Metab. 37 (4) (2022) 674–683.
[40] G.S. Rosenberg, A. Cina, G.R. Schiró, P.D. Giorgi, B. Gueorguiev, M. Alini,
P. Varga, F. Galbusera, E. Gallazzi, Artificial intelligence accurately detects
traumatic thoracolumbar fractures on sagittal radiographs, Medicina 58 (8)
(2022) 998.
[41] N. Hong, S.W. Cho, S. Shin, S.A. Jang, S. Roh, Y. Rhee, H. Kim, K.M. Kim,
S.R. Cummings, Deep learning-based algorithms to detect vertebral fractures and
osteoporosis using lateral spine X-ray radiograph, Bone Rep. 16 (2022) 101576.

11
L.-W. Cheng et al. Neurocomputing 566 (2024) 126946

Hsin-Hung Chou is an associate professor at the Depart- Po-Lun Chu received the Bachelor of Science (B.S) from the
ment of Computer Science and Information Engineering, department of Computer Science and Information Engineer-
National Chi Nan University. He received the Ph.D. degree ing, National Cheng Kung University in 2022. His research
in Computer Science and Information Engineering from interests include computer vision and deep learning.
National Taiwan University at 2004. His research inter-
ests include Computer Game, Machine Learning, Algorithm
Design, Cloud Computing, and so on. In recent decade,
Dr. Chou dedicates in developing the machine learning
techniques for computer games. He has received the re-
ward of the special outstanding researcher at 2015 by
Ministry of Science and Technology. His computer Othello
program Mothello has ever won the silver metal at Com-
puter Olympia 2015 held by ICGA (International Computer
Game Association). And his 9X9 computer Go program I-Szu Cheng received the Doctor of Medicine degree (MD)
Wingo has ever won the bronze metal several times at from the department of Medicine, National Cheng Kung
the Computer Game contests held by TAAI (Association of University in 2023. Her research interests include intelligent
Taiwan Artificial Intelligent) and TCGA (Taiwan Computer medicine and analysis of health insurance research database.
Game Association).

Yu-Xuan Cai is currently studying in the Master’s program

of the Department of Computer Science and Information En-
gineering at National Cheng Kung University. Her interests
include deep learning and intelligent healthcare.

Sun-Yuan Hsieh received the Ph.D. degree in computer

science from National Taiwan University, Taipei, Taiwan,
in June 1998. He then served the compulsory two-year
military service. From August 2000 to January 2002, he
was an assistant professor at the Department of Computer
Science and Information Engineering, National Chi Nan
Kuo-Yuan Huang was awarded a M.D. degree in School of
University. In February 2002, he joined the Department
Medicine from Chang-Gung University in Taiwan in 1996;
of Computer Science and Information Engineering, National
an Ph.D. degree in Institute of Clinical Medicine from
Cheng Kung University, and now he is a chair professor.
National Cheng-Kung University in 2011. He is currently an
Dr. Hsieh has an extremely impressive record of research
associate professor in College of Medicine, National Cheng
achievements in areas of algorithms and fault-tolerant com-
Kung University; and served as an attending orthopedic
puting for interconnection networks. His awards include
surgeon at department of Orthopedics at National Cheng
the 2020 ACM Distinguished scientist, 2019 Kwoh Ting
Kung University Hospital. He is professionally specialized in
Li Honorable Scholar Award, 2016 Outstanding Research
Spine, General Orthopedics, Scoliosis and Osteoporosis. His
Award of Taiwan Ministry of Science and Technology,
research interest focuses on bio-molecular science, biome-
President’s Citation Award (American Biographical Institute)
chanics of intervertebral disc and spine, bone metabolism,
in 2007, Engineering Professor Award of Chinese Institute
joint disease, joint replacement, orthopedic related re-
of Engineers (Kaohsiung Branch) in 2008, National Sci-
searches, osteoporosis, etc. Dr. Huang is a board member of
ence Council’s Outstanding Research Awards in 2009, IEEE
the Pacific and Asian Society of Minimally Invasive Spine
Outstanding Technical Achievement Award (IEEE Tainan
Surgery, active member of the Taiwan Spine Society, active
Section) in 2011, Outstanding Electronic Engineering Pro-
member of the Taiwan orthopedic association, active mem-
fessor Award of Chinese Institute of Electrical Engineers
ber of Formosan Medical Association, a Certified Clinical
in 2013, and Outstanding Engineering Professor Award of
Densitometrist, active Member of the Formosa Association,
Chinese Institute of Engineers in 2014. He is Fellow of the
active member of Regenerative Medicine, active member of
British Computer Society (BCS) and Fellow of Institution of
Taiwan Orthopaedic Research Society, and an active and
Engineering and Technology (IET).
life member of the Asia Pacific Orthopaedic Association
Dr. Hsieh is currently an experienced editor with edito-
Membership.
rial services to a number of journals, including serving as
Chin-Chiang Hsieh was awarded a M.D. in School of associate editors of IEEE Transactions on Computers, IEEE
Medicine, Yang Ming College of Medicine (now College of Transactions on Reliability, IEEE ACCESS, Journal of Com-
Medicine, National Yang Ming Chiao Tung University) in puter and System Science (Elsevier), Theoretical Computer
1992. He is currently a Clinical Instructor in College of Science (Elsevier), Discrete Applied Mathematics (Elsevier),
Medicine, National Cheng Kung University; and served as Journal of Supercomputing (Springer), Editor-in-Chiefs of
an attending radiologist and Director at department of Ra- International Journal of Computer Mathematics (Taylor &
diology in Tainan Hospital, Ministry of Health and Welfare. Francis Group), Parallel Processing Letters (World Scien-
He is professionally specialized in Orthopedic Radiology, tific), Discrete Mathematics, Algorithms and Applications
Genitourinary Radiology, Neuroradiology and Interventional (World Scientific), and Managing editor of Journal of Inter-
Radiology, interpretation of Computed Tomography and connection Networks (World Scientific). In addition, he has
Magnetic Resonance Imaging. His research interest focuses served on organization committee and/or program commit-
on spine stability, traumatic spine injury, burst fracture tee of several dozens international conferences in computer
and vertebral osteonecrosis, dual energy application on science and computer engineering. His current research
fracture and spine instrumentation and imaging of failed interests include design and analysis of algorithms, fault-
back surgery, correlation of bone mineral density with tolerant computing, bioinformatics, parallel and distributed
computed tomography, etc. Dr. Hsieh is a board member of computing, and algorithmic graph theory.
Taiwan Radiological Society and Neuroradiological Society
of Taiwan.

12