TYPE Original Research

PUBLISHED 15 February 2024

DOI 10.3389/fvets.2024.1334438

Detection of spondylosis
OPEN ACCESS deformans in thoracolumbar and
lumbar lateral X-ray images of
EDITED BY
Ryou Tanaka,
Tokyo University of Agriculture and
Technology, Japan

REVIEWED BY
dogs using a deep learning
Seungjo Park,
Chonnam National University,
Republic of Korea
network
Sang-Kwon Lee,
Kyungpook National University,
Republic of Korea Junseol Park1,2† , Hyunwoo Cho3† , Yewon Ji1 , Kichang Lee1 and
*CORRESPONDENCE Hakyoung Yoon1,2*
Hakyoung Yoon
1
[email protected] Department of Veterinary Medical Imaging, College of Veterinary Medicine, Jeonbuk National
University, Iksan, Republic of Korea, 2 Biosafety Research Institute and College of Veterinary Medicine,
†
These authors have contributed equally to Jeonbuk National University, Iksan, Republic of Korea, 3 Department of Electronic Engineering, Sogang
this work and share first authorship University, Seoul, Republic of Korea

RECEIVED 07 November 2023

ACCEPTED 30 January 2024
PUBLISHED 15 February 2024 Introduction: Spondylosis deformans is a non-inflammatory osteophytic
CITATION reaction that develops to re-establish the stability of weakened joints between
Park J, Cho H, Ji Y, Lee K and Yoon H (2024) intervertebral discs. However, assessing these changes using radiography is
Detection of spondylosis deformans in
thoracolumbar and lumbar lateral X-ray
subjective and difficult. In human medicine, attempts have been made to use
images of dogs using a deep learning artificial intelligence to accurately diagnose difficult and ambiguous diseases
network. Front. Vet. Sci. 11:1334438. in medical imaging. Deep learning, a form of artificial intelligence, is most
doi: 10.3389/fvets.2024.1334438
commonly used in medical imaging data analysis. It is a technique that utilizes
COPYRIGHT neural networks to self-learn and extract features from data to diagnose diseases.
© 2024 Park, Cho, Ji, Lee and Yoon. This is an
open-access article distributed under the However, no deep learning model has been developed to detect vertebral
terms of the Creative Commons Attribution diseases in canine thoracolumbar and lumbar lateral X-ray images. Therefore,
License (CC BY). The use, distribution or this study aimed to establish a segmentation model that automatically recognizes
reproduction in other forums is permitted,
provided the original author(s) and the the vertebral body and spondylosis deformans in the thoracolumbar and lumbar
copyright owner(s) are credited and that the lateral radiographs of dogs.
original publication in this journal is cited, in
accordance with accepted academic practice. Methods: A total of 265 thoracolumbar and lumbar lateral radiographic images
No use, distribution or reproduction is from 162 dogs were used to develop and evaluate the deep learning model based
permitted which does not comply with these on the attention U-Net algorithm to segment the vertebral body and detect
terms.
spondylosis deformans.
Results: When comparing the ability of the deep learning model and veterinary
clinicians to recognize spondylosis deformans in the test dataset, the kappa value
was 0.839, indicating an almost perfect agreement.
Conclusions: The deep learning model developed in this study is expected
to automatically detect spondylosis deformans on thoracolumbar and lumbar
lateral radiographs of dogs, helping to quickly and accurately identify unstable
intervertebral disc space sites. Furthermore, the segmentation model developed
in this study is expected to be useful for developing models that automatically
recognize various vertebral and disc diseases.

KEYWORDS

disc instability, intervertebral disc space, artificial intelligence, bony spur, intervertebral
disc disease, canine

Frontiers in Veterinary Science 01 frontiersin.org

Park et al. 10.3389/fvets.2024.1334438

1 Introduction 2 Materials and methods

Spondylolysis deformans is a non-inflammatory degenerative 2.1 Patient dataset
change characterized by new bone formation in the endplates of
the vertebral body (1). It is relatively common in dogs and can This retrospective study included patients who presented to
occur because of disc degeneration and corrected disc instability Jeonbuk National University Animal Medical Center between April
(2, 3). The exact mechanism remains unclear, but it is thought 2017 and October 2023 and underwent radiographic imaging of
to be primarily due to the age-related destruction of peripheral the thoracolumbar or lumbar vertebrae. In 162 dogs, X-ray images
annulus fibers (1, 3, 4). This can lead to discontinuity and (ECO-BT-525 VET, EcoRay, Gwangju, Korea) were obtained and
weakening of disc attachment, stressing the ventral and dorsal used to develop the deep learning models. Patients with and
longitudinal ligaments, which in turn can lead to herniation without specific clinical signs of disc disease were randomly
of the ventral or dorsal disc and development of spondylosis selected and included in the study. For 152 of the 162 dogs,
deformans (3). Previous studies have shown a correlation between we conducted physical and neurologic examinations related to
intervertebral disc protrusion and the location of spondylosis thoracolumbar and lumbar disc disease. We also investigated the
deformans (2, 3, 5, 6). In addition, spondylosis deformans occurs relationship between spondylosis deformans and clinical signs
on the dorsal aspect of the vertebrae and may progress to nerve root associated with thoracolumbar and lumbar disc disease. Dogs were
impingement or cause meningeal irritation, resulting in neurologic considered to have clinical signs if any of the following were
dysfunction (3). identified: proprioceptive ataxia of the pelvic limbs, spinal pain,
Radiologically, spondylosis deformans is characterized by loss of the panniculus reflex, and loss of deep pain. This study was
osteophyte formation in the vertebral endplate and can range approved by the Institutional Animal Care and Use Committee of
from small bony proliferations of the vertebral endplate to Jeonbuk National University (approval nos. JBNU NON2022-085
the bridging of the adjacent vertebrae (1, 6, 7). In cases and NON2023-023).
of obvious lesions, such as spondylosis deformans that form
bridges, the diagnosis can be easily made on radiographs alone;
however, if the lesions are very mild, the diagnosis can be 2.2 Image dataset
somewhat subjective.
The use of deep learning models, a form of artificial intelligence, 2.2.1 Radiographic image acquisition for deep
for the objective and accurate diagnosis of various disc and learning model development
vertebral diseases is an increasingly active area of research in Spondylosis deformans typically initiates as a ventral or dorsal
medicine and has recently shown promising results in various bony proliferation and extends adjacent to intervertebral discs
medical image analysis tasks such as classification, object detection, to fill the gap. Consequently, it is more prominently visible on
and segmentation (8–12). For image classification tasks, deep lateral X-ray images than on ventrodorsal or dorsoventral X-ray
learning models can accurately detect diseases and abnormalities; images. Hence, only lateral X-ray images were used in this study
however, they cannot accurately localize the exact regions of (1, 19). A total of 265 lateral thoracolumbar and lumbar lateral
interest (ROI) (13, 14). Although object detection models can X-ray images (ECO-BT-525 VET; EcoRay, Gwangju, Korea) from
localize rough ROIs, semantic segmentation models offer the 162 dogs were used for model development. For the training and
most accurate localization of abnormalities through pixel-wise validation datasets, we used images acquired under conditions of
classification (8, 15). Accurate localization of the ROI provided 66 kVp−70 kVp and 2.6 mAs−3.0 mAs. In thoracolumbar lateral
by segmentation models can be used to automatically measure X-ray images, the beam center was located at T12-T13 with a
parameters or detect abnormalities with interpretable results (16– field of view (FOV) ranging from approximately T8 to L4, varying
18). In a recent medical study involving humans, a model was slightly between dogs. For lumbar X-ray images, the beam center
developed to automatically detect spondylolisthesis by accurately was located at L3-L4, covering images from T12 to the cranial level
segmenting the lumbar spine in X-ray images (9). However, the of caudal vertebrae. All training and validation dataset images had
use of deep learning models to diagnose vertebral diseases using the FOV set to best exclude abdominal organs, and these images
the X-ray images of dogs has not been studied. In this study, were captured by keeping the vertebrae as straight as possible.
we aimed to develop a novel deep learning-based segmentation The X-ray images used for the test dataset were plane abdominal
model for thoracolumbar and lumbar lateral X-ray images of lateral images acquired at 61 kVp−80 kVp and 10 mAs−14 mAs;
dogs for automatically recognizing vertebral bodies and detecting these conditions depended on the size of the dog, as depicted
spondylosis deformans. Unlike previous studies, we focused on by abdomen width measurements for each dog. The beam center
not only detecting the presence of spondylosis deformans but was located around the last rib, the FOV was set to include all
also accurately segmenting very small lesions, such as grade 1 abdominal organs from the liver to the hips, and the images were
spondylosis deformans, to large lesions, such as grade 3 spondylosis cropped to the same area as the FOV of the training and validation
deformans. We also aimed to design a segmentation model that datasets. All digital radiographic images were post-processed to
accurately separates the vertebral body and spondylosis deformans maintain adequate contrast. Images containing motion artifacts
region with high accuracy by labeling each region accordingly pixel and rotation of vertebrae in the acquired X-ray images were
by pixel. excluded from the study. The dataset distribution for training,

Frontiers in Veterinary Science 02 frontiersin.org

Park et al. 10.3389/fvets.2024.1334438

FIGURE 1
Grading of spondylosis deformans. Grade 0: no bony spur at the vertebral endplate; Grade 1: small bony spur at the edge of the vertebral endplate
not extending past the endplate; Grade 2: more developed bony spurs, but not connected to adjacent vertebra; Grade 3: bony spur connected to
adjacent vertebra forming bony bridges.

FIGURE 2
Example of manual segmentations. In the thoracolumbar lateral X-ray images (A), the vertebral body (yellow), intervertebral disc space (green),
intervertebral foramen (pink), and spondylosis deformans (orange) are labeled with separate colors using a segmentation tool (MediLabel software) to
distinguish them (B).

validation, and test followed an approximate ratio of 80:10:10, and 2.2.2 Evaluation of radiographic images
the training and validation data were chosen randomly. A total On thoracolumbar lateral radiographs, the vertebral body from
of 119 thoracolumbar lateral X-ray images (92 images for training T10 to L3 was evaluated, while on lumbar lateral radiographs, the
data, 13 images for validation data, and 14 images for test data) and vertebral body from L1 to L7 was evaluated. Spondylosis deformans
146 lumbar lateral X-ray images (124 images for training data, 12 was classified into three grades, as shown in Figure 1 (1, 6, 20). The
images for validation data, and 10 images for test data) were used area where spondylosis deformans occurred was checked, and the
to develop deep learning model. most frequently affected area was evaluated.

Frontiers in Veterinary Science 03 frontiersin.org

Park et al. 10.3389/fvets.2024.1334438

FIGURE 3
Example of data augmentation during the training phase. (A) The original image is randomly augmented by (B) adaptive histogram equalization, (C)
horizontal flip, (D) vertical flip, (E) rotation, (F) zoom-in, and (G) zoom-out. All spatial augmentations were performed using cropping or
zero-padding to match the original dimensions. The rotation angle was randomly selected between −45 and 45 and the zoom range was selected
between 0.5× and 1.5× of the original dimensions.

2.3 Deep learning model development resolutions. To maintain the aspect ratio and detailed information
of the images, we set the input resolution to 1024 pixels in
2.3.1 Manual segmentation height and 512 pixels in width. To increase data diversity and
The X-ray images used in this study were manually labeled robustness, the images were augmented by flipping (horizontal,
by 13 veterinary clinicians (residents in the Veterinary Medical vertical), rotation, adaptive histogram equalization (21), zooming
Imaging Department of the Teaching Hospital of Jeonbuk in with random cropping, and zooming out with a zero pad.
National University) using MediLabel software (Ingradient, Inc., Each augmentation was utilized during the training phase with a
Seoul, South Korea). In thoracolumbar and lumbar lateral X- probability of 0.5. Figure 3 shows an example of the augmented
ray images, separate colors were used for labeling to distinguish samples. In addition, the image intensity was normalized into the
spondylosis deformans, vertebral bodies, intervertebral disc space, range of 0 to 1 as shown in Eq. 1, where I is the original image, and
and intervertebral foramen. To label the areas of spondylosis Imax and Imin represent the maximum and minimum values of the
deformans, two veterinarians analyzed the radiographs and selected image intensity, respectively.
areas of common agreement. Figure 2 shows an example of manual
segmentation of a lumbar lateral X-ray image. I − Imin
Inorm = (1)
Imax − Imin

2.3.2 Data preprocessing

Owing to the varying resolutions of the dataset used in this 2.3.3 Network architecture
study, all images were resized to a uniform dimension, and the In this study, a convolutional neural network (CNN)
subsequent predictions were interpolated back to their original segmentation network was employed (Figure 4). The architecture

Frontiers in Veterinary Science 04 frontiersin.org

Park et al. 10.3389/fvets.2024.1334438

FIGURE 4
Structural diagram of the employed deep learning model. In the encoding stage, the features were downsampled (indicated by the orange arrow),
whereas the feature dimensions were expanded using convolution blocks (highlighted by the green arrow). Each convolution block comprised a 3 ×
3 convolution and an activation function. In the decoding stage, the features from the encoding stage were concatenated with the upsampled
features after the implementation of a skip connection and an attention gate. Subsequently, the features were sequentially upsampled to their
original resolutions. The final prediction included three channels, each representing a different pixel class.

FIGURE 5
Examples of manual vs. automated segmentation of vertebrae, intervertebral disc space, and foramen. Results of manual segmentation (A, C) of
vertebral body, spondylosis deformans, intervertebral disc space and intervertebral foramen; automatic segmentation (B, D) in thoracolumbar and
lumbar lateral radiographs.

of this model is based on attention U-Net (22), which has shown 2.3.4 Loss function and implementation details
promising results in medical image segmentation tasks. To cope Recent studies proposed various loss functions for successful
with the large variance in dog size (range: 1.64 kg−36 kg), the segmentation under various conditions (23, 24). To successfully
modified attention U-Net in this study was designed to have train the designed model to simultaneously segment the
deeper feature extraction (i.e., multiscale features) than the original vertebral bodies and detect spondylosis deformans, we utilized
attention U-Net (22) architecture. The designed network extracts a combination of two loss functions. The utilized loss function
features at 7 levels, reducing the spatial resolution from (1024, 512) consists of a region-based loss function and a distribution-based
to (8, 16) for height and width, respectively. The filter dimensions loss function (23, 24). The region-based loss function allows the
of the model (F1 , F2 , F3 , F4 , F5 , F6 , F7 ) were selected as deep learning model to successfully segment the vertebral bodies
16, 32, 64, 128, 256, 512, and 1024, respectively. The attention from images, and the objective of the distribution-based loss
gate introduced in the attention U-Net retained in its original function is to create a deep learning model to detect spondylosis
structure (22). deformans pixel-wise. The weighted categorical cross-entropy loss

Frontiers in Veterinary Science 05 frontiersin.org

Park et al. 10.3389/fvets.2024.1334438

TABLE 1 Cohen’s kappa analysis between a veterinary clinician and deep learning model for detection of spondylosis deformans.

Kappa value Sensitivity (%) Specificity (%) PPV (%) NPV (%)
∗
Validation dataset 0.813 87.5 98.5 77.8 99.3
∗
Test dataset 0.839 98.4 86.3 86.1 98.4
95% CI, 95% confidence interval; PPV, positive predictive value; NPV, negative predictive value.
Experimental values were considered significant at p < 0.01∗ .

TABLE 2 Detection rate based on the grade of spondylosis deformans by

agreement for the vertebral body, intervertebral disc space, and
deep learning.
foramen. The DSC is a relative measure of the percentage of pixels
Grade Sensitivity No. of detected Total no. of that overlap between auto-segmentation and manual segmentation
(%) spondylosis spondylosis images (30). The closer the DSC is to 1, the better the match
deformans by deformans between the two segmentations. DSC was performed using the
deep learning
following equation (30):
1 96.2 25 26

2 100 26 26

3 100 11 11 DSC = 2 (Intersected region)/(sum of region segmentations)

(25) and focal Tversky loss function (26) were used. The total loss To evaluate how closely matched spondylosis deformans were
function is expressed as follows: detected by the deep learning model and veterinary clinicians,
Cohen’s kappa analysis was performed to check sensitivity and
1 + yŷ
TI(y, ŷ) = specificity. Additionally, Cohen’s kappa test was used to determine
1 + yŷ + β(1 − y)ŷ + (1 − β)y(1 − ŷ) inter-veterinarian agreement on the areas determined to have
X
3
spondylosis deformans. Cohen’s kappa results were interpreted
LFTL (ŷ, y) = (1 − TIc )γ
as follow: values ≤ 0.00–0.20 indicated non- to slight, 0.21–0.40
c=1
indicated fair, 0.40–0.60 indicated moderate, 0.60–0.80 indicated
X
3
LWCE (ŷ, y) = − w(c) × y(c) log ŷ(c) substantial, and 0.80–1.00 indicated almost perfect agreement (31).
c=1 To identify the relationship between the presence of
LTOTAL (ŷ, y) = LWCE (ŷ, y) + LFTL (ŷ, y) spondylosis deformans and the occurrence of thoracolumbar
disc disease-related neurologic signs, a Chi-square test was
where y, ŷ represent the prediction and ground-truth probability performed. The values were determined to be statistically
map, c is the corresponding class (i.e., background, normal significant at p < 0.05.
vertebral body, spondylosis deformans), w is the class weighting SPSS version 29.0 (SPSS Corp., Armonk, NY, USA) was used for
factor, and β, γ are hyperparameters. The w was set to 1, 1, and statistical analyses.
5 for the background, normal vertebral body, and spondylosis
deformans. β, γ were set at 0.3 and 0.75. The weighting factors
and hyperparameters were determined empirically to obtain the
best results. To optimize the utilized loss function, the Adam 3 Results
optimiser (27) with a learning rate of 1e−4 was used, and early
stopping criteria were used to obtain the best result. The proposed 3.1 Animals
methods were implemented using MONAI (28) and PyTorch
(29) frameworks. A total of 29 breeds were enrolled in the study: Maltese (n = 39),
Pomeranian (n = 18), Poodle (n = 16), Dachshund (n = 13), mixed
breed (n = 11), Shih Tzu (n = 9), Pekingese (n = 9), Cocker Spaniel
2.4 Time measurement (n = 8), Miniature Poodle (n = 5), Chihuahua (n = 4), Bichon Frise
(n = 4), Beagle (n = 3), Yorkshire Terrier (n = 3), Golden Retriever
For the 25 lateral thoracolumbar and lumbar X-ray images used (n = 2), German Shepherd (n = 2), Old English Sheepdog (n = 2),
as validation data, the time required per image to detect spondylosis Jindo (n = 2), Boston Terrier (n = 1), Labrador Retriever (n = 1),
deformans by a veterinary clinician and by the deep learning model Miniature Pinscher (n = 1), Pompitz (n = 1), Samoyed (n = 1),
through auto-segmentation was recorded and compared. Siberian Husky (n = 1), Schnauzer (n = 1), Shetland Sheepdog (n
= 1), Spitz (n = 1), Welsh Corgi (n = 1), Sapsaree (n = 1), and
Whippet (n = 1). The average weight was 7.42 kg (range: 1.64–
2.5 Model accuracy and statistical analysis 36 kg), and the average age was 8.81 years (range: 0.7–17 years), and
for seven dogs, there were no information regarding body weight.
Dice similarity coefficient (DSC) was used to confirm whether This study included 91 male dogs (24 intact, 67 castrated) and 71
auto-segmentation and manual segmentation were in close female dogs (20 intact, 51 spayed).

Frontiers in Veterinary Science 06 frontiersin.org

Park et al. 10.3389/fvets.2024.1334438

TABLE 3 Relationship of the presence or grade of spondylosis and clinical signs related to the disc disease in 152 dogs.

Grade 0 Grade 1 Grade 2 Grade 3 Total

No. of dogs with neurologic signs 32 (21.1%) 11 (7.2%) 8 (5.3%) 7 (4.6%) 58 (38.2%)

No. of dogs not presenting neurologic signs 64 (42.1%) 13 (8.6%) 9 (5.9%) 8 (5.3%) 94 (61.8%)

Total 96 (63.1%) 24 (15.8%) 17 (11.2%) 15 (9.9%) 152 (100%)

TABLE 4 Number of spondylsis deformans cases and grading of dogs in

3.2 Deep learning model shows 119 thoracolumbar lateral X-ray images.
considerable similarity compared to
manual segmentation in recognizing T10- T11- T12- T13- L1- L2-
T11 T12 T13 L1 L2 L3
vertebral body
Grade 1 0 (0) 5 (0) 11 (0) 8 (0) 6 (2) 7 (0)

To evaluate the performance of the model, the DSC between Grade 2 2 (0) 3 (1) 6 (0) 1 (0) 6 (1) 7 (0)
manual and automated segmentation was calculated for a Grade 3 0 (0) 0 (0) 4 (1) 2 (0) 6 (1) 3 (1)
validation dataset of 25 dogs. Post-processing, the average DSC
T, thoracic; L, lumbar; Numbers in parentheses indicate number of dorsal spondylosis
value for the vertebral body was 0.910 ± 0.038 (mean ± SD). deformans cases.
The DSC values of the intervertebral disc space and foramen
TABLE 5 Number of spondylsis deformans cases and grading of dogs in
were 0.787 ± 0.056 and 0.779 ± 0.083, respectively. Figure 5
146 lumbar lateral X-ray images.
shows an example of manual and automated segmentation of the
vertebral body, spondylosis deformans, intervertebral disc space, L1- L2- L3- L4- L5- L6-
and intervertebral foramen. L2 L3 L4 L5 L6 L7
Grade 1 11 (3) 11 (0) 5 (1) 6 (3) 1 (1) 2 (2)

Grade 2 5 (1) 7 (0) 4 (2) 7 (3) 2 (1) 2 (1)

Grade 3 5 (0) 5 (2) 4 (2) 4 (1) 3 (0) 2 (0)

3.3 Deep learning model shows a high L, lumbar; Numbers in parentheses indicate number of dorsal spondylosis deformans cases.
capability of detecting spondylosis
deformans in a short time
The mean time taken by deep learning to automatically segment
The interclass correlation kappa value between veterinary vertebral bodies and detect spondylosis deformans was found to be
clinicians for the evaluation of spondylosis deformans in 0.052 s per image, while the time taken by a veterinary clinician to
thoracolumbar and lumbar lateral X-ray images was 0.889, evaluate spondylosis deformans was found to be 17.92 s per image.
confirming almost perfect agreement. In addition, when the
deep learning model recognized spondylosis deformans on
thoracolumbar and lumbar lateral X-ray images in validation 3.4 Spondylosis deformans was more
dataset, the kappa value with identification by a veterinary clinician common at T12-T13 and L2-L3 in
was 0.813, indicating an almost perfect agreement. The sensitivity thoracolumbar and lumbar X-ray images
was 87.5%, specificity was 98.5%, positive predictive rate was
77.8%, and negative predictive rate was 99.3% (Table 1). Upon A total of 265 thoracolumbar and lumbar lateral radiographs
checking the ability of the deep learning model on abdominal of dogs were evaluated for spondylosis deformans. Of these, 163
lateral X-ray images in the test dataset, the kappa value was 0.839, dogs exhibited spondylosis deformans (thoracolumbar vertebrae:
indicating an almost perfect agreement. The sensitivity was 98.4 77 sites, lumbar vertebrae: 86 sites), of which 30 (18.4%) had dorsal
%, specificity was 86.3%, positive predictive rate was 86.1%, and spondylosis deformans (Tables 4, 5). Within the thoracolumbar
negative predictive rate was 98.4% (Table 1). When we checked how vertebrae (T10-L3), the most commonly affected sites included
well deep learning recognized 63 sites with spondylosis deformans T12-T13 (36.7%) (Figure 6). In the lumbar vertebrae (L1-L7), the
identified by veterinary clinicians in the test dataset, it detected most commonly affected sites were L2-L3 (26.7%) (Figure 7). In
spondylosis deformans in 96.2% of the cases in grade 1 and 100% contrast, T10-T11 (2.6%), L5-L6 (7.0%), and L6-L7 (7.0%) were
of the cases in grades 2 and 3 (Table 2). identified as the least affected areas.
The relationship between spondylosis deformans and clinical
signs associated with thoracolumbar and lumbar disc disease
was examed in 152 dogs. If a dog displayed multiple grades of 4 Discussion
spondylosis deformans, we selected the highest grade lesion for
comparison. In this study, ∼33.3% of dogs without spondylosis This study is the first to develop deep learning models to
deformans, 45.8% with grade 1, 47.1% with grade 2, and 46.7% with automatically segment the vertebral body and detect spondylosis
grade 3 exhibited clinical signs associated with thoracolumbar and deformans in the thoracolumbar and lumbar lateral radiographs
lumbar disc disease (Table 3). of dogs. Recently, models to detect specific diseases in X-ray

Frontiers in Veterinary Science 07 frontiersin.org

Park et al. 10.3389/fvets.2024.1334438

FIGURE 6
Distribution of sites for spondylosis deformans and grading of dogs in 119 thoracolumbar lateral X-ray images. T, thoracic; L, lumbar.

FIGURE 7
Distribution of sites for spondylosis deformans and grading of dogs in 146 lumbar lateral X-ray images. L, lumbar.

images of dogs have been developed and commercialized, and However, in this study, the DSC for the intervertebral disc
studies have been conducted to auto-segment gross head and space and intervertebral foramen were slightly lower, reaching up to
neck tumors for radiation therapy and automatically detect kidney 0.787 and 0.779 respectively. The DSC value for the intervertebral
calculi and measure kidney volume in CT images in dogs; however, disc space and foramen were unavailable in the aforementioned
there have been no studies in the literature on a deep learning human study, which is likely because the size of the disc space and
model to automatically recognize disc diseases in X-ray images of intervertebral foramen is relatively smaller than that of the vertebral
dogs (32–34). bodies. Inconsistent imaging of the vertebral endplates can cause
First, a segmentation model was developed to automatically disc space distances to appear shorter than they actually are or
detect the vertebrae and disc space and was then compared to deep results in unclear margins. Moreover, intervertebral foramen may
learning models developed for lumbar vertebrae detection in prior have ambiguous borders with adjacent structures. Additionally,
human literature. The mean intersection over union value, a term the substantial relative weight difference between breeds in dogs
similar to DSC for vertebral bodies, was found to be 0.8–0.88 in a compared to humans might have also affected this value. We
human study, and when the DSC value of the vertebral body (0.910) assume that training with a larger number of images will be
in this study was converted to the mean intersection over union required to improve the recognition rate of disc spaces and
value, it was found to be 0.835, showing the ability to recognize intervertebral foramen.
the vertebral body at a similar level to the previous human In this study, the most affected areas of spondylosis deformans
research (9). in thoracolumbar and lumbar lateral radiographs of dogs were

Frontiers in Veterinary Science 08 frontiersin.org

Park et al. 10.3389/fvets.2024.1334438

identified as T12-T13 and L2-L3, with T12-L5 being more affected When the association between the presence of spondylosis
overall. Previous studies (3, 4, 35) have found that the most deformans and thoracolumbar and lumbar disc disease-related
affected area was somewhat different (L1-L3); however, the overall clinical signs was examined, no statistical significance was found.
commonality of occurrence in the vertebral region (T12-L5) was However, dogs with spondylosis deformans were more likely to
similarly confirmed. exhibit clinical signs associated with disc disease. However, no
Of the total 163 cases of spondylosis deformans identified association was found between clinical signs and the grade of
in 265 animals, 30 cases of dorsal spondylosis deformans spondylosis deformans, and the presence of spondylosis deformans
(18.4%) were identified. However, of the total 63 spondylosis did not differ by breed in this study. Previous studies also
deformans in the test dataset, only 3 cases of dorsal spondylosis have similarly shown that spondylosis deformans is detected on
deformans (2 sites for grade 1 and 1 site for grade 2) were radiographic images in 17.8%−32.8% of normal dogs without
identified, but the deep learning model accurately recognized specific clinical signs related to disc diseases (35–37). However
all of them as spondylosis deformans. Although the number a higher rate of prevalence of spondylosis deformans (∼47%) is
of images with dorsal spondylosis deformans in the test identified in patients with intervertebral disc protrusion, a form of
dataset is insufficient to evaluate accuracy, in the training and intervertebral disc disease (IVDD) (38). Hence, while the presence
validation datasets, the deep learning model also demonstrated of spondylosis deformans is not specific for IVDD, it can be useful
accurate recognition in the training and validation datasets, in the diagnosis of disc disease when considered in conjunction
identifying 26 out of 27 dorsal spondylosis deformans. This with radiologic findings such as intervertebral space narrowing.
suggests a high likelihood of accurate recognition for dorsal In conclusion, the deep learning model developed in this
spondylosis deformans. study is expected to help clinical veterinarians accurately and
Even for very small lesions such as grade 1 spondylosis rapidly evaluate spondylosis deformans by determining sites of
deformans, the model successfully detected 25 out of 26 disc instability. Moreover, the vertebral body segmentation model
sites, achieving a remarkably high sensitivity of 96.2% and a developed in this study can be applied to develop deep learning
very high negative predictive value of 98.4%. The specificity models that automatically detect vertebral body diseases such as
and positive predictive values were also high, at 86.3% and tumors, discospondylitis, dislocation of vertebrae, and abnormal
86.1%, respectively. These results suggest that the deep learning narrowed intervertebral disc space.
model in this study exhibited higher sensitivity toward very
small spondylosis deformans than two clinicians. This may
likely be attributed to the criteria of the model, which that Data availability statement
considered a site as true spondylosis deformans only if both
clinicians identified it as spondylosis deformans. Notably, when The raw data supporting the conclusions of this article will be
we identified 10 sites in our test dataset as spondylosis made available by the authors, without undue reservation.
deformans using the deep learning model alone, 6 of them
were sites judged as spondylosis deformans by one of the
two clinicians. Ethics statement
In addition, the time taken by the deep learning model
to auto-segment spondylosis deformans in radiographic images The animal studies were approved by the Institutional Animal
was significantly faster than that by a veterinary clinician to Care and Use Committee of Jeonbuk National University (approval
visually evaluate them, which we believe could ultimately save the nos. JBNU NON2022-085 and NON2023-023). The studies were
interpretation time of veterinary clinicians. conducted in accordance with the local legislation and institutional
In case of the test dataset, images of the vertebral region requirements. Written informed consent was obtained from the
cropped from the lateral abdominal X-ray were used, when owners for the participation of their animals in this study.
checked ability of the deep learning model to detect spondylosis
deformans, observed that the sensitivity, specificity, positive
predictive rate, and negative predictive rate were all higher Author contributions
than 86%. These results mean that the deep learning model
can detect spondylosis deformans highly accurately not only in JP: Conceptualization, Data curation, Investigation, Writing
thoracolumbar and lumbar lateral X-ray images taken with precise – original draft, Formal analysis, Methodology, Validation,
vertebral endplate alignment but also in lateral abdominal X-ray Writing – review & editing. HC: Conceptualization, Investigation,
images taken routinely for general medical examinations, which Methodology, Writing – original draft, Writing – review
is expected to be useful in veterinary clinical practice. However, & editing, Data curation, Formal analysis, Validation. YJ:
when applying the deep learning model to uncropped plane Conceptualization, Methodology, Writing – review & editing,
lateral abdominal X-ray images, detection capability of vertebral Investigation. KL: Conceptualization, Project administration,
bodies and spondylosis deformans is slightly lower than applying Writing – review & editing, Supervision. HY: Conceptualization,
cropped images. Therefore, additional advancements training are Data curation, Investigation, Project administration, Supervision,
considered necessary to fully apply it to general plane lateral Writing – review & editing, Methodology, Validation, Writing –
abdominal X-ray images. original draft.

Frontiers in Veterinary Science 09 frontiersin.org

Park et al. 10.3389/fvets.2024.1334438

Funding Conflict of interest

The author(s) declare financial support was received for The authors declare that the research was conducted in the
the research, authorship, and/or publication of this article. This absence of any commercial or financial relationships that could be
study was supported by the National Research Foundation of construed as a potential conflict of interest.
Korea and funded by a grant from the Korean Government
(no. 2021R1C1C1006794).
Publisher’s note
Acknowledgments All claims expressed in this article are solely those of the
authors and do not necessarily represent those of their affiliated
The authors would like to thank the veterinary clinicians of the organizations, or those of the publisher, the editors and the
Veterinary Medical Imaging Department of the Teaching Hospital reviewers. Any product that may be evaluated in this article, or
of Jeonbuk National University for their assistance with the manual claim that may be made by its manufacturer, is not guaranteed or
segmentation of the X-ray images used in this study. endorsed by the publisher.

References
1. Thomas WB, Fingeroth JM. Spondylosis deformans. Adv Int Disc Dis Dogs and 16. Cho H, Song I, Jang J, Yoo Y. A lightweight deep learning network on a system-
Cats. 2015:67–74. doi: 10.1002/9781118940372.ch8 on-chip for wearable ultrasound bladder volume measurement systems: preliminary
study. Bioengineering. (2023) 10:525. doi: 10.3390/bioengineering10050525
2. Morgan JP, Miyabayashi T. Degenerative changes in the vertebral column of the
dog. Vet Radiol. (1988) 29:72–7. doi: 10.1111/j.1740-8261.1988.tb01752.x 17. Cho H, Kim D, Chang S, Kang J, Yoo Y. A system-on-chip solution for
deep learning-based automatic fetal biometric measurement. Exp Syst Appl. (2024)
3. Levine GJ, Levine JM, Walker MA, Pool RR, Fosgate GT. Evaluation of the
12:121482. doi: 10.1016/j.eswa.2023.121482
association between spondylosis deformans and clinical signs of intervertebral disk
disease in dogs: 172 cases (1999–2000). J Am Vet Med Assoc. (2006) 228:96– 18. Mansoor A, Bagci U, Foster B, Xu Z, Papadakis GZ, Folio LR, et al. Segmentation
100. doi: 10.2460/javma.228.1.96 and image analysis of abnormal lungs at CT: current approaches, challenges, and future
trends. Radiographics. (2015) 35:1056–76. doi: 10.1148/rg.2015140232
4. Morgan J, Hansson K, Miyabayashi T. Spondylosis deformans in the
female beagle dog: a radiographic study. J Small Anim Prac. (1989) 30:457– 19. Flegel T. Spinal Radiographs for Neurological Diseases. Advancing the Veterinary
60. doi: 10.1111/j.1748-5827.1989.tb01607.x Profession in Eastern Europe. Berlin: Humboldt-University (2019), 179.
5. da Costa RC, De Decker S, Lewis MJ, Volk H. Diagnostic imaging in intervertebral 20. Halle KS, Granhus A. Veterinary chiropractic treatment as a measure
disc disease. Front Vet Sci. (2020) 7:588338. doi: 10.3389/fvets.2020.588338 to prevent the occurrence of spondylosis in boxers. Vet Sci. (2021)
8:199. doi: 10.3390/vetsci8090199
6. Kranenburg HC, Voorhout G, Grinwis GCM, Hazewinkel HAW, Meij
BP. Diffuse idiopathic skeletal hyperostosis (DISH) and spondylosis deformans 21. Zuiderveld K. Contrast limited adaptive histogram equalization. Graphics Gems.
in purebred dogs: a retrospective radiographic study. Vet J. (2011) 190:e84– (1994) 12:474–85. doi: 10.1016/B978-0-12-336156-1.50061-6
90. doi: 10.1016/j.tvjl.2011.04.008
22. Oktay O, Schlemper J, Folgoc LL, Lee M, Heinrich M, Misawa K, et al. Attention
7. Togni A, Kranenburg HJ, Morgan JP, Steffen F. Radiographic and MRI u-net: Learning where to look for the pancreas. arXiv [Preprint]. arXiv:180403999
characteristics of lumbar disseminated idiopathic spinal hyperostosis and spondylosis (2018).
deformans in dogs. J Small Anim Pract. (2014) 55:343–9. doi: 10.1111/jsap.
23. Jadon S. A Survey of Loss Functions for Semantic Segmentation. In: 2020 IEEE
12218
Conference on Computational Intelligence in Bioinformatics and Computational Biology
8. Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F, Ghafoorian M, et al. A (CIBCB). Piscataway, NJ: IEEE (2020).
survey on deep learning in medical image analysis. Med Image Anal. (2017) 42:60–
24. Ma J. Segmentation loss odyssey. arXiv [Preprint]. arXiv:200513449 (2020).
88. doi: 10.1016/j.media.2017.07.005
25. Pihur V, Datta S, Datta S. Weighted rank aggregation of cluster validation
9. Trinh GM, Shao HC, Hsieh KL, Lee CY, Liu HW, Lai CW, et al. Detection of
measures: a Monte Carlo cross-entropy approach. Bioinformatics. (2007) 23:1607–
lumbar spondylolisthesis from X-ray images using deep learning network. J Clin Med.
15. doi: 10.1093/bioinformatics/btm158
(2022) 11:450. doi: 10.3390/jcm11185450
26. Abraham N, Khan NM. A Novel Focal Tversky Loss Function With Improved
10. Merali Z, Wang JZ, Badhiwala JH, Witiw CD, Wilson JR, Fehlings MG, et al. deep
Attention u-net for Lesion Segmentation. In: 2019 IEEE 16th International Symposium
learning model for detection of cervical spinal cord compression in MRI scans. Sci Rep.
on Biomedical Imaging (ISBI 2019). Piscataway, NJ: IEEE (2019).
(2021) 11:10473. doi: 10.1038/s41598-021-89848-3
27. Adam KDBJ. A method for stochastic optimization. arXiv [Preprint].
11. Kijowski R, Liu F, Caliva F, Pedoia V. Deep learning for lesion detection,
arXiv:14126980 (2014) 1412.
progression, and prediction of musculoskeletal disease. J Magn Reson Imaging. (2020)
52:1607–19. doi: 10.1002/jmri.27001 28. Cardoso MJ, Li W, Brown R, Ma N, Kerfoot E, Wang Y, et al. Monai: An open-
source framework for deep learning in healthcare. arXiv [Preprint]. arXiv:221102701
12. Zheng HD, Sun YL, Kong DW, Yin MC, Chen J, Lin YP, et al. Deep learning-
(2022).
based high-accuracy quantitation for lumbar intervertebral disc degeneration from
MRI. Nat Commun. (2022) 13:841. doi: 10.1038/s41467-022-28387-5 29. Paszke A, Gross S, Massa F, Lerer A, Bradbury J, Chanan G, et al. Pytorch: An
imperative style, high-performance deep learning library. Adv Neural Inf Proc Syst.
13. Cai L, Gao J, Zhao D. A review of the application of deep learning
(2019) 32.
in medical image classification and segmentation. Annal Transl Med. (2020)
8:44. doi: 10.21037/atm.2020.02.44 30. Zou KH, Warfield SK, Bharatha A, Tempany CM, Kaus MR, Haker SJ, et al.
Statistical validation of image segmentation quality based on a spatial overlap index.
14. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-
Acad Radiol. (2004) 11:178–89. doi: 10.1016/S1076-6332(03)00671-8
cam: Visual Explanations From Deep Networks via Gradient-Based Localization. In:
Proceedings of the IEEE International Conference on Computer Vision. Piscataway, NJ: 31. Munoz SR, Bangdiwala SI. Interpretation of kappa and B statistics measures of
IEEE (2017). agreement. J Appl Stat. (1997) 24:105–12. doi: 10.1080/02664769723918
15. Yang R, Yu Y. Artificial convolutional neural network in object detection 32. Park J, Choi B, Ko J, Chun J, Park I, Lee J, et al. Deep-learning-based automatic
and semantic segmentation for medical imaging analysis. Front Oncol. (2021) segmentation of head and neck organs for radiation therapy in dogs. Front Vet Sci.
11:638182. doi: 10.3389/fonc.2021.638182 (2021) 8:721612. doi: 10.3389/fvets.2021.721612

Frontiers in Veterinary Science 10 frontiersin.org

Park et al. 10.3389/fvets.2024.1334438

33. Ji Y, Cho H, Seon S, Lee K, Yoon H, A. deep learning model for CT-based kidney 36. Morgan JP. Spondylosis deformans in the dog: a morphologic study with
volume determination in dogs and normal reference definition. Front Vet Sci. (2022) some clinical and experimental observations. Acta Orthop Scand. (1967) 38:1–
9:1011804. doi: 10.3389/fvets.2022.1011804 88. doi: 10.3109/ort.1967.38.suppl-96.01
34. Ji Y, Hwang G, Lee SJ, Lee K, Yoon H. A deep learning model for automated 37. Read RM, Smith RN. A comparison of spondylosis deformans in the English
kidney calculi detection on non-contrast computed tomography scans in dogs. Front and Swedish cat and in the English dog. J Small Anim Prac. (1968) 9:159–
Vet Sci. (2023) 10: 1236579. doi: 10.3389/fvets.2023.1236579 66. doi: 10.1111/j.1748-5827.1968.tb04605.x
35. Morgan JP, Ljunggren G, Read R. Spondylosis deformans (vertebral 38. Macias C, McKee W, May C, Innes J. Thoracolumbar disc disease
osteophytosis) in the dog: a radiographic study from England, Sweden and USA. in large dogs: a study of 99 cases. J Small Anim Prac. (2002) 43:439–
J Small Anim Prac. (1967) 8:57–66. doi: 10.1111/j.1748-5827.1967.tb04526.x 46. doi: 10.1111/j.1748-5827.2002.tb00010.x

Frontiers in Veterinary Science 11 frontiersin.org