IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 13, No. 3, September 2024, pp. 3414∼3425

ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i3.pp3414-3425 ❒ 3414

Computer aided detection for vertebral deformities

diagnosis based on deep learning
Nabila Ounasser1 , Maryem Rhanoui2,3 , Mounia Mikram3 , Bouchra El Asri1
1 IMS Team, ADMIR Laboratory, Rabat IT Center, ENSIAS, Mohammed V University, Rabat, Morocco
2 Laboratory Health, Systemic, Process, Research Unit 4129, University Claude Bernard Lyon 1, Lyon, France
3 Meridian Team, LYRICA Laboratory, School of Information Sciences, Rabat, Morocco

Article Info ABSTRACT

Article history: The diagnosis of spinal deformities is one of the most frequent daily clinical
routine. X-ray images are used to diagnose several pathologies in order to re-
Received Sep 25, 2023
duce harmful radiations of the patient. Spinal deformities are diagnosed es-
Revised Feb 26, 2024 sentially from vertebral shapes, orientations, and positions, so their detection
Accepted Mar 11, 2024 and segmentation are major steps required for diagnosis. Deep learning could
be applied for automatic diagnosis to detect scoliosis and its variants with a
Keywords: favourable performance. In this study, based on 609 spinal anterior-posterior
x-ray images obtained from the public SpineWeb, we examine generative ad-
Automatic spine diagnosis versarial network (GAN) based architectures and convolutional neural network
Convolutional neural network (CNN) based architectures models that are capable of automatically detecting
Deep learning anomalies in radiograph and achieve expert-level performances in various fields
Generative adverserial network providing a solid comparative study. Most of the implemented models are apt
Medical imaging to automatically distinguish limits between vertebrae so determining their shape
Scoliosis with a very good visual performance. The GAN-based architecture estimates the
Spinal deformity required vertebral landmarks with an accuracy rate of 0.966, signify its capacity
for automatic scoliosis assessment in a clinical setting.

This is an open access article under the CC BY-SA license.

Corresponding Author:
Nabila Ounasser
IMS Team, ADMIR Laboratory, Rabat IT Center, ENSIAS, Mohammed V University
Rabat, Morocco
Email: [email protected]

1. INTRODUCTION
Musculoskeletal irregularities represent the most common medical condition, leading to persistent dis-
comfort and impairment over time. Consequently, accurately identifying abnormalities in radiographic images
is an essential undertaking in the field of medicine [1]. Evaluating X-rays to diagnose orthopedic ailments such
as bone deformities, tumors, and fractures is a labor-intensive process that demands the expertise of qualified
professionals. Consequently, the creation of a computer-assisted diagnostic system for detecting anomalies in
X-ray images has garnered significant interest [2].
The spine is the pillar of the body, it is the substrate of the musculoskeletal system that is breathable
of our mobility, it supports and sustains the body and the structure of its organs. Despite their criticality, spinal
pathologies are often unaware of the diagnosis, especially spinal deformities. Spinal deformity is an abnormal
alignment or curve of the bony vertebral column. Early detection and orthotic treatment of scoliosis would
reduce the need for surgical intervention [3]. Therefore, computer-assisted assistance is needed for an efficient
and early detection of these pathologies, allowing an effective prevention or treatment [4].

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938 ❒ 3415

The current state of the art in the field of vertebral deformity diagnosis has seen a significant shift
towards the utilization of deep learning techniques. These advancements have revolutionized the accuracy and
efficiency of computer-aided detection systems, enabling more precise identification and characterization of
vertebral abnormalities. Recent studies have demonstrated the effectiveness of deep learning models in auto-
matically detecting and classifying various types of vertebral deformities from medical imaging data, offering
a promising avenue for improving clinical diagnosis and patient care. Computer vision is introduced for im-
age analysis due to its promising performance in extracting information from images. Many tasks have been
performed by computer vision, including automated anomaly detection [2], identification and classification of
fracture [5], [6], diabetic retinopathy screening [7], and skin lesion classification [8]. Several deep learning
models have been investigated in this direction such as generative adversarial networks (GANs) [9]–[11] and
convolutional neural network (CNN) [12] that facilitate anomalies detection and achieved expert-level per-
formances in various fields. Most of these approaches focus on computed tomography (CT) datasets only.
However, these methods are rarely applicable to X-ray images because of additional difficulties. Radiography
is used for the diagnosis of various pathologies. It allows the visualization of a change in volume or a struc-
tural abnormality. The cross-sectional images obtained allow to evaluate the shape, position, volume, size, and
possible abnormalities of a multitude of anatomical structures, depending on the region being explored. Also
X-ray images have a lower resolution. All these facts are detrimental to the automation of detection procedures
for X-ray data sets.
Spinal image processes are poorly seen on radiographic images, which is common, frequent, and re-
mains the first reflex in clinical practice. In addition, transverse processes are usually not seen at all because
they are outside the acquisition volume. Therefore, we focus on detecting the vertebral bodies and then delin-
eating the entire vertebrae. Additionally, GANs and CNNs are being forcefully explored for anomaly detection
[2], [5], [6] in other areas including intrusion detection [13], fraud detection [14], to protect valuable systems,
and since there is no more valuable that our human body, in this study we will consider the human body as our
system and protect it from anomalies.
In this research, we aim to investigate both approaches with the goal of applying them in the field of
medicine, specifically for a comprehensive comparative study on the detection of spinal deformities, particu-
larly scoliosis, at varying degrees. Our study focuses on enhancing the quality of spinal deformity detection in
X-ray image tasks. We aim to address this need by leveraging the power of GANs and CNNs to improve the ac-
curacy and efficiency of detecting orthopedic irregularities in radiographic images. To achieve this, we illustrate
the utilization of GANs and CNNs for identifying orthopedic irregularities in radiographic images. We validate
their effectiveness in detecting scoliosis at different severity levels using a publicly available dataset comprising
609 anterior-posterior spinal X-ray images sourced from SpineWeb (http://spineweb.digitalimaginggroup.ca).
The subsequent sections of this paper are structured as follows: we begin with the background section,
followed by section 3, which provides a concise overview of related research. Section 4 outlines the method-
ology employed in our study, while section 5 details the materials used, including the dataset and implemented
models. Section 6 delves into the discussion of the results, and to finally conclude our research in the section 7.

2. RELATED WORK
X-ray analysis is a widely used medical method for diagnosing orthopedic conditions, including bone
deformities, tumors, and fractures. In this section, we conduct a review of existing deep learning models
developed for detecting anomalies in orthopedic musculoskeletal radiographs. Numerous researchers have
trained CNNs on bone X-ray images. Dias [15] employed transfer learning techniques such as feature extraction
and fine-tuning to enhance the detection of musculoskeletal abnormalities in X-ray images.
Recent research has explored deep anomaly detection methods, such as GANs like AlphaGAN [16],
BiGAN [17], and more to improve anomaly detection tasks. Researchers have made efforts to enhance the
performance of these models by modifying their components. For instance, [18], [19] GANomaly has seen im-
provements through extensions like skip-connections which employ an autoencoder to map the reconstructed
input back to the latent space. Song et al. [20] proposed a Res-unetGAN model based on the GAN architecture
and applied it to Mura. This network consists of two parts: a generator and a discriminator. The encoder com-
ponent of the generator employs ResNet50 to extract features from normal samples and obtain their potential
feature vector representations.
Lately, there has been renewed interest in spine detection and spinal shape analysis. Several deep

Computer aided detection for vertebral deformities diagnosis based on deep learning (Nabila Ounasser)
3416 ❒ ISSN: 2252-8938

learning models have been developed for spine-related tasks, utilizing X-rays, MRIs, or CT images. For ex-
ample, Yi et al. [21] has proposed models for spine-related tasks. Additionally, several researchers [3], [22]
have explored spine-related tasks using different imaging modalities. Han et al. [23] introduced SpineGAN,
a model designed to handle the complex and variable nature of spinal structures. SpineGAN incorporates an
atrous convolution autoencoder module to capture semantic task-aware representations while preserving fine-
grained structural information. He et al. [24] propose one-stage methods capable of simultaneously segmenting
discs, vertebrae, and neural foramen using GAN-based models. Deep neural networks have also been em-
ployed to detect spine vertebrae, leading to significant improvements in performance. Du et al. [25] introduced
SpineNet, a backbone architecture with scale-permuted intermediate features and cross-scale connections that
was learned through neural architecture search. Wu et al. [26] introduced an innovative approach for automati-
cally estimating landmarks in adolescent idiopathic scoliosis (AIS) assessment by combining CNN (ConvNet)
with statistical techniques to accommodate the variability seen in X-ray images. More recently, Yeh et al. [27]
tackled the task of automatically detecting landmarks and performing alignment analysis in whole-spine lateral
radiographs, employing a deep learning approach. Cina et al. [28] proposed a trainable two-step deep learning
approach for landmark localization in spine radiographs. Furthermore, Zukić et al. [29] have employed CNNs
for detecting vertebra centers.
In Table 1, we present a summary of recent methods employed in the field of spine deformity detection.
These methods encompass a range of approaches, including CNNs, autoencoder architectures, and traditional
machine learning techniques. While these existing methods have made significant strides in spine deformity
detection, our study aims to introduce novel advancements to further enhance the accuracy and efficiency of
diagnosis. Building upon the foundations laid by previous research, we propose the integration of GAN into the
diagnostic pipeline. By harnessing the power of GANs for synthetic data generation and feature representation
learning, we anticipate a substantial improvement in the detection of subtle spine abnormalities and variations.
Through rigorous experimentation and validation, we expect our proposed methodology to outperform existing
approaches, offering clinicians a more reliable and comprehensive tool for early diagnosis and personalized
treatment planning.

Table 1. Summary of recent works of spine deformity detection

[Ref] Dataset Approach
[3] EOS imaging system Introduce an automated method for extracting anatomical parameters from biplanar ra-
diographs of the spine.
[21] ASCE MICCAI 2019 Introduce a method for accurately detecting landmarks in AIS, crucial for precise Cobb
challenge angle estimation. By localizing vertebra centers and tracing corner landmarks through
learned offsets.
[25] ILSVRC-2012 and COCO Propose Spinenet to optimize performance by training a backbone network to efficiently
datasets handle scale variations, thereby enhancing recognition and localization accuracy.
[27] Clinical dataset Presents a deep learning approach for automatically detecting landmarks and analyzing
alignment in whole-spine lateral radiographs. The proposed method aims to identify land-
marks and assess alignment in spinal images.
[28] IRCCS Istituto Ortopedico Introduce a 2-step deep learning model tailored for landmark localization in spine radio-
Galeazzi graphs. The approach aims to enhance accuracy in identifying key anatomical landmarks
crucial for diagnostic assessments.
[29] Clinical datasets Propose a robust detection and segmentation method for diagnosing vertebral diseases
using routine MRI images. The approach aims to detect and segment vertebral abnormal-
ities, facilitating more precise diagnosis and treatment planning.

3. BACKGROUND
3.1. Vertebra detection
Detecting vertebrae involves employing various methods, each with its own set of techniques and
algorithms. First, one prominent technique is the utilization of the Viola-Jones method [29]. This method
primarily focuses on detecting the centers of the vertebrae, providing a foundational step in the overall process.
The second approach involves employing an object detector to identify the vertebrae as bounding box entities.
These bounding box objects are subsequently inputted into a landmark regression network as distinct images.
The integration of both methods, alongside advancements in deep learning techniques, has significantly en-
hanced the accuracy and efficiency of vertebrae detection systems. Figure 1 illustrates this process, showcasing

Int J Artif Intell, Vol. 13, No. 3, September 2024: 3414–3425

Int J Artif Intell ISSN: 2252-8938 ❒ 3417

the transformation from bounding box detection to landmark-based reconstruction on the original images.

Figure 1. Pipeline for spine detection

3.2. Scoliosis diagnosis

Orthopedic anomalies are frequent reasons for consultation from childhood. Several pathologies at-
tack the structure of the spine (vertebral fracture, inflammation of the discs, deformation of the spine) to detect
these phathologies, doctors may need multimodal radiologies (magnetic resonance imaging (MRI), CT scans,
and X-ray) depending on the type of disease. In our study, we focus on spinal deformity types which can be
diagnosed from vertebra detection using X-ray images as the primary material.
Scoliosis with its varying degrees is a deformation of the spine in the 3 planes of space. It is typically
identified during childhood or the early teenage years. The spine normally exhibits natural curves in the cervi-
cal, thoracic, and lumbar regions, aligning in the ”sagittal” plane. These inherent curves serve to align the head
with the pelvis and act as shock absorbers, evenly distributing mechanical stresses during bodily movement.
Scoliosis, however, is commonly described as an abnormal curvature of the spine in the ”coronal” (frontal)
plane. Despite its measurement primarily occurring in the frontal plane, scoliosis is, in fact, a more intricate
condition.

4. METHODS
To achieve our objective, we implemented state-of-the-art GANs and CNNs, most famous families
in deep learning, Figure 2, tailored to the unique challenges of spinal deformity detection. Our approach
involved fine-tuning the models using different techniques notably data augmentation techniques to effectively
capture the intricate features indicative of various deformity types. This adaptation process was crucial, as
existing models were not specifically designed for this mission. We chose GANs for their potential to generate
synthetic data, which we anticipated would enhance the models’ performance in detecting subtle deformities.
In this section we will review CNN and GAN models investigated in this study. Those families of models apply
the second method of Vertebra detection as explained in the background section.

Figure 2. A summary of the concepts encompassing artificial intelligence

Computer aided detection for vertebral deformities diagnosis based on deep learning (Nabila Ounasser)
3418 ❒ ISSN: 2252-8938

4.1. Overview of generative adversarial networks

GANs [30] is considered as one of the most powerful member of the neural network family, due to re-
alistic data-generation capacities. GANs offer a significant advantage in their capacity to generate data, which
has led to their successful application in various computer vision tasks such as anomaly detection, image gener-
ation, and image super-resolution [2]. In musculoskeletal imaging, automating the detection and segmentation
of vertebral degenerative disease is crucial for expediting and streamlining the radiology diagnostic workflow.
Deep learning methods have been extensively employed in this field, including GAN-based approaches, which
are particularly adaptable. As showen in Figure 3 GANs operate by training two competing networks: a gen-
erator and a discriminator. The generator produces realistic synthetic samples from noise (the z-latent space),
while the discriminator discerns between genuine and synthetic samples. This flexible architecture has been
utilized for tasks like identifying the location of vertebrae, discs, and spinal shape.

Figure 3. GAN’s architecture

4.1.1. SpineGAN
SpineGAN [23] was developed to detect spinal abnormalities and uncovering potential underlying
pathological factors. The architecture of SpineGAN consists of two networks Figure 4, each comprising three
modules. Firstly, there is a specialized segmentation network tasked with segmenting and classifying neural
foramen, intervertebral discs, and vertebrae in radiological images. This segmentation network integrates a
deep atrous convolution autoencoder module for encoding spinal images and conducting pixel-level classifi-
cation. Additionally, it incorporates a recurrent neural network (RNN) module based on local long short term
memory network (LSTM) to dynamically model the spatial relationships among different spinal structures in
pathology. In alignment with the principles of GANs, a discriminative network is introduced to oversee and
motivate the segmentation network, ensuring the generation of accurate predictions.

Figure 4. SpineGAN’s architecture [23]

4.1.2. Randomized generative adversarial network

Randomized generative adversarial network (RandGAN) [31] for COVID-19 detection. Its architec-
ture composed of two components Figure 5: generator and a discriminator. The particularity of RandGAN’s
architecture is the Inception and residual block. To enhance the generalizability of RandGAN’s generator,

Int J Artif Intell, Vol. 13, No. 3, September 2024: 3414–3425

Int J Artif Intell ISSN: 2252-8938 ❒ 3419

random images are drawn from the training class cohort and encoded using inception layers. This approach
offers variability in both random noise vectors and real image representations during generator training. The
inception and residual architecture aims to improve GAN’s ability to capture fine details and maintain spatial
information across convolution and pooling layers. However, increasing the generator’s depth for capturing
distant details, while theoretically valid, poses stability and training challenges for deep GANs.

Figure 5. RandGAN’s generator architecture [31]

4.1.3. CycleGAN
CycleGAN is one of the first models to have attracted a lot of attention through image-to-image trans-
lation using unpaired images [22]. It is composed of two generators and two discriminators as shown in
Figure 6. The first generator transforms X into Y and the second one transforms Y into X. The first discrim-
inator have to differentiate real images sampled from X and images produced by the second generator, then
this generator is updated accordingly to get a better performance. The second discriminator attempts to differ-
entiate real images sampled from Y and images produced by the first generator, then this generator is updated
accordingly to get a better performance. That’s what we call the competitive learning, it is a technique focused
on improving the model’s performance.

Figure 6. CycleGAN’s architecture

4.2. Overview of convolutional neural networks

CNNs, a category of artificial neural networks that have gained prominence in various computer vision
tasks, are now garnering attention in diverse domains, including radiology. Detecting anomalies in medical

Computer aided detection for vertebral deformities diagnosis based on deep learning (Nabila Ounasser)
3420 ❒ ISSN: 2252-8938

images is a common challenge for radiologists, as these anomalies are infrequent and must be identified amidst
numerous normal cases. Recent radiomics studies have explored traditional machine learning models, including
techniques for feature extraction, image analysis, and object detection. As we see in Figure 7 a CNN consists
of multiple stacked convolutional layers, each with the ability to recognize increasingly complex patterns.

Figure 7. CNN’s architecture

This sequential design enables CNNs to learn hierarchical features. Importantly, CNNs do not nec-
essarily require human expert segmentation of anomalies. Through dimensionality reduction, CNNs extract
features from images and transform them into a lower-dimensional representation while retaining essential in-
formation. In contrast, other deep learning approaches tend to be more computationally intensive, necessitating
the use of graphical processing units (GPUs) for model training.

4.2.1. ResNet50
ResNet introduces a residual learning framework designed to facilitate the training of deeper neural
networks. Its distinguishing feature lies in the establishment of connections between numerous layers, which
simplifies the optimization of the underlying residual mapping, denoted as H(x). There are several variations
of the ResNet network model, which differ based on the number of convolutional layers they incorporate. In
our particular case, we have chosen to employ the ResNet50 variant, which boasts a depth of 50 layers.

4.2.2. BoostNet
The BoostNet architecture [26] is crafted for the automatic detection of spinal landmarks to facili-
tate a comprehensive assessment of AIS. The BoostNet architecture effectively addresses the limitations of
traditional AIS assessments by enhancing the feature space through the removal of outliers and bolstering
robustness by enforcing the integrity of the spinal structure. This architecture comprises three fundamental
components. First, a set of convolutional layers serves as feature extractors, autonomously learning features
from the dataset. Second, a BoostLayer is employed to eliminate the influence of detrimental outlier features.
Finally, a spinal structured multi-output layer functions as a prior mechanism to mitigate the impact of a limited
dataset, capturing crucial relationships between each spinal landmark.

4.2.3. SpineNet
SpineNet represents a CNN backbone distinguished by its scale-permuted intermediate features and
cross-scale connections, a structure acquired through the process of neural architecture search during training
for object detection tasks. This innovative architecture was crafted based on scale-permuted models and was
intentionally designed for a fair comparison with ResNet Figure 8. Du et al. [25] introduced four distinct ar-
chitectures within the SpineNet family, each excelling in various latency-performance trade-offs, thus offering
versatility for a wide range of use cases. The models are denoted as SpineNet-49/96/143/190. The difference
is the feature dimensions in the entire network and number of blocs that constitute the model.

Int J Artif Intell, Vol. 13, No. 3, September 2024: 3414–3425

Int J Artif Intell ISSN: 2252-8938 ❒ 3421

Figure 8. Spine’s architecture (scale-permuted model) [25]

5. EXPERIMENT
5.1. Dataset
The dataset contains 609 anterior-posterior radiographic images of the spine obtained from the public
SpineWeb repository (http://spineweb.digitalimaginggroup.ca). All images show varying degrees of scoliosis
symptoms. They manually annotated the landmarks, each image contains 68 GT landmarks corresponding to
the 4 corners of the 17 vertebrae, and 3 Cobb angles. Datatest contains images without GT. During training,
the landmarks were scaled to the dimensions of the original image, so that the range of values belonging to the
interval [0,1] depends on the location of the landmark relative to the original image. 80% of the dataset was for
training (487 images) and 20% for testing (122 images) no patient is placed in both sets. The project code and
resources utilized in this study are publicly available on GitHub at: https://github.com/nabinabila/Vertebral-
Deformities-Diagnosis-based-on-Deep-Learning.

5.2. Preprocessing
During the preprocessing stage, we implemented data augmentation. This involved enhancing the im-
ages to introduce greater diversity into the dataset. This augmentation was performed to enable the models to
acquire a deeper understanding of the dataset by learning high-level features that remain consistent despite typ-
ical affine transformations, such as horizontal flips, which might occur when generating radiographic images.

6. RESULTS AND DISCUSSION

The results of our implemented models compare favorably with results presented in previous works.
Upon reviewing the accuracy of the previously implemented models, it’s evident that their performance falls
within the range of 0.520 percent to 0.966 percent. Numerous factors can influence the models’ performance,
including architectural approach, layer design, padding, shape, normalization, activation, loss function, opti-
mizer, batch size, learning rate, pooling, and output layer. Achieving an effective outcome was our primary
objective after extensive tuning efforts. Many of our models featured multiple layers and modules, which typi-
cally impose a substantial computational burden. Training these models sometimes extended over several days
and running them on basic hardware or standard laptop configurations proved to be excessively time-consuming
for the dataset.
Furthermore, preprocessing is one of the key for good results in data science tasks. After choosing the
deep learning model for the study, it is necessary to prepare a large amount of data. Image size is a parameter
that impacts the accuracy of detecting the boundaries between vertebrae. In our experiment, we obtained an
acceptable result with the image resolution of 256 × 256 pixels. We explored data augmentation to cope with
the limited data available to us, which increases the amount of data in the training phase. The drawbacks of
this system that we need to pay attention to are the rotation methods, excessive compression, and shear, as
they can impact the performance of intervertebral disc boundary detection. Detailly, Table 2 shows that Spine-
GAN, CycleGAN, and RandGAN on average achieves the best accuracy (0.966; 0.922; 0.913 percent). This
demonstrates effectiveness of the GAN-based architectures, their modules that are capable to get a deep and
accurate representation by conserving the differences between normal and anomalous structures.

Computer aided detection for vertebral deformities diagnosis based on deep learning (Nabila Ounasser)
3422 ❒ ISSN: 2252-8938

Table 2. Detection results on X-Ray images

CNN-based architecture GAN-based architecture
Evaluation Metrics ResNet50 ConvNet BoostNet SpineNet49 SpineNet143 SpineGAN CycleGAN RandGAN
Accuracy 0.520 0.563 0.917 0.875 0.933 0.966 0.922 0.913
MSE 0.026 0.018 0.006 0.0057 0.0051 0.0046 0.0052 0.0077
Precision 0.459 0.438 0.877 0.866 0,890 0.981 0.933 0.903
Detection Speed 1.26 1.43 4.12 8.56 9.12 7.21 6.33 8.11

While with less processing time the CNNs approaches, Convnet and Resnet50 run faster than GANs,
they have lower rate of performance and do not provide orientation estimates. SpineNet models achieved an ac-
curacy of 0.875 and 0.933 percent. In particular, the largest model, SpineNet-143, outperform by 0.933 percent
wich is an impressive result for a single model without multi-scale testing during inference. BoostNet attained
a commendable accuracy of 0.917 percent, primarily attributable to the contributions of the BoostLayer and the
spinal structured multi-output regression layer. These components effectively captured the structural details
of the spinal landmark coordinates. Moreover, our models exhibited impressive precision values, reflecting
their ability to correctly identify true positive cases while minimizing false positive detections. The high accu-
racy and precision achieved by our GAN models underscores their reliability and robustness in detecting spine
deformities, instilling confidence in their clinical utility and potential for real-world deployment.
In addition to accuracy and precision, remarkably, our GAN models achieved consistently low mean
squared error (MSE) values, indicating their proficiency in accurately estimating the extent of deformations
and their spatial distribution within the spine images. This fine-grained analysis is invaluable for clinicians
in evaluating the severity and progression of spinal abnormalities, facilitating personalized treatment planning
and monitoring. Furthermore, our GAN models demonstrated impressive detection speed, enabling rapid and
efficient analysis of large volumes of spine imaging data. Leveraging parallel computing architectures and
optimized model architectures, our models achieved near-real-time performance without compromising accu-
racy. This high-speed processing capability enhances the scalability and practicality of our approach, making
it well-suited for integration into clinical workflows and telemedicine applications.
Visually in Figures 9 and 10, the illustration serves as a qualitative showcase of GAN’s proficiency in
detecting spinal landmarks. Regardless of differences in anatomy and image contrast among various patients,
GAN consistently and accurately identifies all spinal landmarks. It’s noteworthy that the landmarks detected
by GAN exhibit a closer conformity to the spinal shape when compared to the performance of ConvNet.
In comparison to previous studies utilizing CNNs for spine deformity detection, our GAN-based ap-
proach demonstrated notable advancements in both accuracy and robustness. While CNNs have been widely
adopted in medical image analysis due to their ability to automatically extract hierarchical features, they of-
ten struggle with capturing subtle deformities and variations in spine images. In contrast, our GAN models
leverage adversarial training to generate synthetic data, effectively augmenting the training set and enhancing
the models’ ability to generalize across diverse deformity patterns. As a result, our GAN models consistently
outperformed CNN-based approaches in detecting spine deformities, achieving higher area under the curve
(AUC) scores, precision values, and lower MSE.
Similarly, our findings surpass those reported in studies employing autoencoder architectures for spine
deformity detection. Although autoencoder models excel in unsupervised feature learning and data compres-
sion, they may struggle with preserving important anatomical details and discriminating between normal and
abnormal spine configurations. In contrast, our GAN models leverage the discriminative power of adversarial
training to explicitly learn the underlying features indicative of spine deformities, thereby achieving superior
performance in terms of both accuracy and clinical relevance. By integrating both generative and discriminative
components, our GAN models strike a balance between data generation and discrimination, resulting in more
effective and interpretable representations of spine deformities.
Furthermore, our study extends beyond the limitations of previous approaches by incorporating a
comprehensive evaluation of detection speed, an aspect often overlooked in existing literature. While CNN and
autoencoder models have demonstrated promising results in terms of accuracy, their computational efficiency
and real-time performance remain areas of concern. In contrast, our GAN models exhibit impressive detection
speed, enabling rapid analysis of spine images without compromising accuracy. This improvement in speed-
to-accuracy ratio is particularly significant in clinical settings, where timely diagnosis and treatment are critical
for patient care. Our study represents a significant advancement in the field of computer-aided detection for
vertebral deformities, showcasing the potential of deep learning techniques in improving diagnostic accuracy

Int J Artif Intell, Vol. 13, No. 3, September 2024: 3414–3425

Int J Artif Intell ISSN: 2252-8938 ❒ 3423

and efficiency. To our knowledge, this is the first study to examine several models for automatic detection
for diagnosis of spinal deformity using X-Ray images. Our observations, from this comparative study, those
methods are an effective way to improve orthopedic anomalies detection tasks. In summary, our comparative
analysis highlights the superior performance of GAN models in spine deformity detection compared to previous
studies utilizing CNN and autoencoder architectures. These findings offer a promising tool for early detection
in spinal deformities.

Figure 9. Examples of landmarks detection on X-rays: Convnet

Figure 10. Examples of landmarks detection on X-rays: GAN

7. CONCLUSION
In this paper, we applied CNN and GAN models, most powerful members of the neural network
family. Unfortunately they are not explored to diagnosis spinal pathologies. Although the spine is the pillar
of the body, it is the substrate of the musculoskeletal system that is breathable of our mobility it supports
and sustains the body and the structure of its organs. There is not enough studies that invest to improve
medical process for this organ. So, our goal was to examined those models for spinal disease analysis. We had
compared and analysed several GAN-based architectures and CNN-based architectures for spinal deformities
detection. Summing up the results, it can be concluded that the deep learning methods here presented were
apt to automatically determine the spine shape with a very good visual performance. We believe that those
methods provide great assistance to clinical experts in orthopedic process analysis. With the improvement
of those methods, they will have the potential to be the key for an automated radiological analysis of spinal
pathologies, in condition of availability a large training dataset. To conclude, this experiment allowed us to
identify the limitations of the models. Future work will explore ways to present a novel approach that could
learn specific features for identifying musculoskeletal abnormalities.

Computer aided detection for vertebral deformities diagnosis based on deep learning (Nabila Ounasser)
3424 ❒ ISSN: 2252-8938

REFERENCES
[1] S. Gyftopoulos, D. Lin, F. Knoll, A. M. Doshi, T. C. Rodrigues, and M. P. Recht, “Artificial intelligence in musculoskeletal imaging:
current status and future directions,” American Journal of Roentgenology, vol. 213, no. 3, pp. 506–513, 2019.
[2] N. Ounasser, M. Rhanoui, M. Mikram, and B. E. Asri, “Generative and autoencoder models for large-scale mutivariate unsupervised
anomaly detection,” Smart Innovation, Systems and Technologies, vol. 237, pp. 45–58, 2022, doi: 10.1007/978-981-16-3637-0 4.
[3] F. Galbusera et al., “Fully automated radiological analysis of spinal disorders and deformities: a deep learning approach,” European
Spine Journal, vol. 28, no. 5, pp. 951–960, 2019, doi: 10.1007/s00586-019-05944-z.
[4] A. S. Rawat, A. Rana, A. Kumar, and A. Bagwari, “Application of multi layer artificial neural network in the diagnosis
system: A systematic review,” IAES International Journal of Artificial Intelligence, vol. 7, no. 3, pp. 138–142, 2018, doi:
10.11591/ijai.v7.i3.pp138-142.
[5] N. Ounasser, M. Rhanoui, M. Mikram, and B. E. Asri, “Anomaly detection in orthopedic musculoskeletal radiographs using deep
learning,” in Proceedings of Eighth International Congress on Information and Communication Technology, 2023, pp. 93–102, doi:
10.1007/978-981-99-3243-6 8.
[6] N. Ounasser, M. Rhanoui, M. Mikram, and B. E. Asri, “Enhancing computer-assisted bone fractures diagnosis in musculoskeletal
radiographs based on generative adversarial networks,” International Journal of Advanced Computer Science and Applications, vol.
14, no. 7, pp. 960–966, 2023, doi: 10.14569/IJACSA.2023.01407104.
[7] G. T. Zago, R. V. Andreão, B. Dorizzi, and E. O. T. Salles, “Diabetic retinopathy detection using red lesion localization and
convolutional neural networks,” Computers in Biology and Medicine, vol. 116, 2020, doi: 10.1016/j.compbiomed.2019.103537.
[8] J. Zhang, Y. Xie, Y. Xia, and C. Shen, “Attention residual learning for skin lesion classification,” IEEE transactions on medical
imaging, vol. 38, no. 9, pp. 2092–2103, 2019, doi: 10.1109/TMI.2019.2893944.
[9] Z. Qin, Z. Liu, P. Zhu, and Y. Xue, “A GAN-based image synthesis method for skin lesion classification,” Computer Methods and
Programs in Biomedicine, vol. 195, 2020, doi: 10.1016/j.cmpb.2020.105568.
[10] Y. Zhou, B. Wang, X. He, S. Cui, and L. Shao, “DR-GAN: conditional generative adversarial network for fine-grained lesion
synthesis on diabetic retinopathy images,” IEEE Journal of Biomedical and Health Informatics, vol. 26, no. 1, pp. 56–66, 2022, doi:
10.1109/JBHI.2020.3045475.
[11] K. B. Park, S. H. Choi, and J. Y. Lee, “M-GAN: retinal blood vessel segmentation by balancing losses through stacked deep fully
convolutional networks,” IEEE Access, vol. 8, pp. 146308–146322, 2020, doi: 10.1109/ACCESS.2020.3015108.
[12] T. Ren et al., “Convolutional neural network detection of axillary lymph node metastasis using standard clinical breast MRI,”
Clinical Breast Cancer, vol. 20, no. 3, pp. e301–e308, 2020, doi: 10.1016/j.clbc.2019.11.009.
[13] S. Huang and K. Lei, “IGAN-IDS: an imbalanced generative adversarial network towards intrusion detection system in ad-hoc
network,” Ad Hoc Networks, vol. 105, 2020, doi: 10.1016/j.adhoc.2020.102177.
[14] U. Fiore, A. D. Santis, F. Perla, P. Zanetti, and F. Palmieri, “Using generative adversarial networks for improving classification
effectiveness in credit card fraud detection,” Information Sciences, vol. 479, pp. 448–455, 2019, doi: 10.1016/j.ins.2017.12.030.
[15] D. D. A. Dias, “Musculoskeletal abnormality detection on x-ray using transfer learning,” Ph.D. Thesis, Department of Intelligent
Interactive Systems, Universitat Pompeu Fabra, Barcelons, Spain, 2019.
[16] K. Raza and N. K. Singh, “A tour of unsupervised deep learning for medical image analysis,” Current Medical Imaging Reviews,
vol. 17, no. 9, pp. 1059–1077, 2022, doi: 10.2174/18756603mtezonzmk0.
[17] J. Donahue, T. Darrell, and P. Krähenbühl, “Adversarial feature learning,” arXiv-Computer Science, 2017.
[18] S. Akcay, A. A. -Abarghouei, and T. P. Breckon, “Skip-GANomaly: skip connected and adversarially trained encoder-
decoder anomaly detection,” in 2019 International Joint Conference on Neural Networks (IJCNN), 2019, pp. 1–8, doi:
10.1109/IJCNN.2019.8851808.
[19] S. Akcay, A. A. -Abarghouei, and T. P. Breckon, “GANomaly: semi-supervised anomaly detection via adversarial training,” in
Computer Vision – ACCV 2018, 2019, pp. 622–637, doi: 10.1007/978-3-030-20893-6 39.
[20] S. Song, K. Yang, A. Wang, S. Zhang, and M. Xia, “A Mura detection model based on unsupervised adversarial learning,” IEEE
Access, vol. 9, pp. 49920–49928, 2021, doi: 10.1109/ACCESS.2021.3069466.
[21] J. Yi, P. Wu, Q. Huang, H. Qu, and D. N. Metaxas, “Vertebra-focused landmark detection for scoliosis assessment,” in 2020 IEEE
17th International Symposium on Biomedical Imaging (ISBI), 2020, pp. 736–740, doi: 10.1109/ISBI45749.2020.9098675.
[22] R. Oulbacha and S. Kadoury, “MRI to CT synthesis of the lumbar spine from a pseudo-3D cycle GAN,” in 2020 IEEE 17th
International Symposium on Biomedical Imaging (ISBI), 2020, pp. 1784–1787, doi: 10.1109/ISBI45749.2020.9098421.
[23] Z. Han, B. Wei, A. Mercado, S. Leung, and S. Li, “Spine-GAN: semantic segmentation of multiple spinal structures,” Medical
Image Analysis, vol. 50, pp. 23–35, 2018, doi: 10.1016/j.media.2018.08.005.
[24] J. He, W. Liu, Y. Wang, X. Ma, and X.-S. Hua, “SpineOne: a one-stage detection framework for degenerative discs and ver-
tebrae,” in 2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), Dec. 2021, pp. 1331–1334, doi:
10.1109/BIBM52615.2021.9669541.
[25] X. Du et al., “Spinenet: learning scale-permuted backbone for recognition and localization,” in 2020 IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR), 2020, pp. 11589–11598, doi: 10.1109/CVPR42600.2020.01161.
[26] H. Wu, C. Bailey, P. Rasoulinejad, and S. Li, “Automatic landmark estimation for adolescent idiopathic scoliosis assessment using
boostnet,” in Medical Image Computing and Computer Assisted Intervention, 2017, pp. 127–135, doi: 10.1007/978-3-319-66182-
7 15.
[27] Y. C. Yeh, C. H. Weng, Y. J. Huang, C. J. Fu, T. T. Tsai, and C. Y. Yeh, “Deep learning approach for automatic landmark detection and
alignment analysis in whole-spine lateral radiographs,” Scientific Reports, vol. 11, no. 1, 2021, doi: 10.1038/s41598-021-87141-x.
[28] A. Cina et al., “2-step deep learning model for landmarks localization in spine radiographs,” Scientific Reports, vol. 11, no. 1, 2021,
doi: 10.1038/s41598-021-89102-w.
[29] D. Zukić, A. Vlasák, J. Egger, D. Hořı́nek, C. Nimsky, and A. Kolb, “Robust detection and segmentation for diagnosis of vertebral
diseases using routine MR images,” Computer Graphics Forum, vol. 33, no. 6, pp. 190–204, 2014, doi: 10.1111/cgf.12343.
[30] I. J. Goodfellow et al., “Generative adversarial nets,” in Advances in Neural Information Processing Systems, 2014, pp. 1–9.
[31] S. Motamed, P. Rogalla, and F. Khalvati, “RANDGAN: randomized generative adversarial network for detection of COVID-19 in
chest X-ray,” Scientific Reports, vol. 11, no. 1, 2021, doi: 10.1038/s41598-021-87994-2.

Int J Artif Intell, Vol. 13, No. 3, September 2024: 3414–3425

Int J Artif Intell ISSN: 2252-8938 ❒ 3425

BIOGRAPHIES OF AUTHORS

Nabila Ounasser holds baccalaureate’s degree in Mathematic science and a state engi-
neering diploma in Data and Knowledge from the School of Information Sciences (ESI). Currently,
she is pursuing a Ph.D. at the ENSIAS (École Nationale Supérieure d’Informatique et d’Analyse des
Systèmes) within the Department of Computer Science. Her research focuses on exploring artificial
intelligence for anomaly detection within the IT architecture and model driven systems development
(IMS) team. Her research areas of interest include artificial intelligent and computer vision. She can
be contacted at email: [email protected].

Bouchra El Asri currently holds the position of Teaching Research Director at ENSIAS
(École Nationale Supérieure d’Informatique et d’Analyse des Systèmes). She was a Technical Direc-
tor at Cyber Machine. She has successfully led two major projects for prominent national organiza-
tions. Additionally, she holds several key roles, including Department Head of Software Engineering,
Coordinator of the Software Engineering program at ENSIAS, and responsibility for the development
of an enhanced version of the program. She is actively involved in the institution as a member of the
governing council, pedagogical committee, and budget monitoring committee. Furthermore, they
played a crucial role in transitioning the Software Engineering program to online teaching during the
COVID-19 crisis. She has a strong research background, having supervised and currently supervising
multiple doctoral theses in the fields of software architecture and data management for healthcare,
industry, and education. Her expertise and contributions extend to scientific committees, doctoral
study centers, and teaching modules within the Software Engineering program at ENSIAS. She can
becontacted at email: [email protected] or [email protected].
Maryem Rhanoui is an Associate Professor specializing in Computer Sciences and Data
Engineering. She received an engineering degree in computer science then a Ph.D. degree from
ENSIAS, Mohammed V University, Rabat 2015. Her research interests include artificial intelligence,
knowledge extraction, and decision making, and medical data analysis. She can be contacted at email:
[email protected].

Mounia Mikram is an Associate Professor of Computer Sciences and Mathematics at the

School of Information Sciences, Rabat since 2010. She received her master degree from Mohammed
V University Rabat (2003) and her Ph.D. degree from Mohammed V University, Rabat, and Bordeaux
I University (2008). Her research interests include pattern recognition, computer vision, biometrics
security systems, and artificial intelligence. She can be contacted at email: [email protected].

Computer aided detection for vertebral deformities diagnosis based on deep learning (Nabila Ounasser)