Osteoporosis, which originates from Greek, is literally

translated as porous bone, [1]. According to World

Health

Organization, low bone mass and microarchitectural

degeneration of bone tissue are the characteristics of

osteoporosis, a progressive systemic skeletal disease that

increases bone fragility and fracture susceptibility, [1].

Osteoporosis “is a metabolic bone condition in which

osteoclastic bone resorption is not counteracted at the cellular

level by osteoblastic bone synthesis. As a result, bones

become brittle and weak, putting them at risk of fracture.

Traditional osteoporosis pathophysiology centered on

endocrine factors such as estrogen shortage or vitamin D

deficiency, as well as secondary hyperparathyroidism.

Although osteoporosis can affect persons of any age or

gender, it is typically an age-related disease that affects

women more than males” [2].

Osteoporosis is diagnosed by dual-energy X-ray

absorptiometry (DXA), which measures bone mineral density

(BMD). Therefore, checking for osteoporosis can have a big

impact on how patients turn out. However, because

osteoporosis is hidden until severe fragility fractures,

osteoporosis is mostly misdiagnosed, and DXA screening for

osteoporosis has been underutilized [3]. Patients frequently

underestimate the severity of the sickness and, as a result,

refuse to volunteer for the screening program [4].

There is a growing consensus that other screening

approaches are needed to overcome the shortcomings of DXA

as an osteoporosis diagnosis method. Adults frequently

undergo Abdominal-Pelvic Computed Tomography (APCT)

to examine a variety of disorders during routine health check-

ups or to follow up on previously identified conditions. Even

if only a tiny percentage of these scans were for osteoporosis

screening opportunistically, there would be a significant

impact. APCT has shown promising results in opportunistic

osteoporosis screening in several studies [5], [6], [9].

Artificial intelligence (AI) and Deep Learning (DL) is

used for image interpretation for osteoporosis classification

[7]. In a 2019 review paper, AI advancements have aided in

the detection of osteoporosis [8]. The following methods were

employed: dental radiographs [9], [13], spine radiographs

[7], [14], hand and wrist radiographs [10], [11], [12], [13].

This study uses a dataset of knee radiographs (i.e.,

knee-Xray images) to apply and compare the training time

of two robust transfer learning model algorithms. The

transfer learning models applied were GoogLeNet, and

VGG-16. In addition, “to compare the diagnostic

performance of the two models, several state-of-the-art

neural networks metric was used.”

In order to estimate the prevalence of osteoporosis in

postmenopausal women, machine learning techniques were

applied [14]. The researchers constructed a non-linear model

using regression support vector machines (SVM) for a sample

of 305 postmenopausal women to ascertain the association

between BMD, diet, and lifestyle variables. A preliminary

Transfer Learning Model Training Time Comparison for Osteoporosis

Classification on Knee Radiograph of RGB and Grayscale Images

1USMAN BELLO ABUBAKAR, 2MOUSSA MAHAMAT BOUKAR

3STEVE ADESHINA, 4SENOL DANE

1Department of Computer Science, Baze University, Abuja, NIGERIA

2Department of Computer Science, Nile University of Nigeria, Abuja, NIGERIA

3Department of Computer Engineering, Nile University of Nigeria Abuja, NIGERIA

4Department of Physiology, Nile University of Nigeria, Abuja, NIGERIA

Abstract: In terms of financial costs and human suffering, osteoporosis poses a serious public health burden.

Reduced bone mass, degeneration of the microarchitecture of bone tissue, and an increased risk of fracture are

its main skeletal symptoms. Osteoporosis is caused not just by low bone mineral density, but also by other

factors such as age, weight, height, and lifestyle. Recent advancement in Artificial Intelligence (AI) has led to

successful applications of expert systems that use Deep Learning techniques for osteoporosis diagnosis based

on some modalities such as dental radiographs amongst others. This study uses a dataset of knee radiographs

(i.e., knee-Xray images) to apply and compare the training time of two robust transfer learning model

algorithms: GoogLeNet, VGG-16, and ResNet50 to classify osteoporosis. The dataset was split into two

subcategories using python opencv library: Grayscale Images and Red Green Blue (RGB) images. From the

scikit learn python analysis, the training time of the GoogLeNet model on grayscale images and RGB images

was 42minutes and 50 minutes respectively. The VGG-16 model training time on grayscale images and RGB

images was 37 minutes and 44 minutes respectively. In addition, to compare the diagnostic performance of the

two models, several state-of-the-art neural networks metric was used.

Keywords: Osteoporosis, Transfer Learning Models, Dual-Energy X-ray Absorptiometry, Bone Mineral

Density

Received: April 29, 2021. Revised: June 22, 2022. Accepted: July 18, 2022. Published: September 13, 2022.

1.Introduction

2. Related Works

WSEAS TRANSACTIONS on ELECTRONICS

DOI: 10.37394/232017.2022.13.7

Usman Bello Abubakar, Moussa Mahamat Boukar,

Steve Adeshina, Senol Dane

E-ISSN: 2415-1513

Volume 13, 2022

Fig. 1. Osteoporosis case and normal case.

The dataset, after statical augmentation using python

augmentation functions, comprises 323 normal knee

radiograph images and 323 osteoporotic knee radiograph

images of patients. Table I shows the splitting of image data

into a train, test and validation data.

TABLE I. IMAGE DISTRIBUTION

Class

Total

Training

Validation

Testing

Normal(0)

323

207

Osteoporosis(1)

323

207

In “most image data, the pixel values are integers with

values ranging from 0 to 255. Since neural networks only

analyze inputs with modest weight values, inputs with large

integer values can interfere with or slow down the learning

process.” Therefore, picture normalization is a recommended

practice: pixel values range between 0 and 1. The images in

the dataset were normalized(rescaled) using the python

ImageDataGenerator method and passing rescale=1. /255 as

its argument.

The two image formats considered in this study are RGB

images and Grayscale images. The dataset consists of images

in RGB format. An RGB (red, green, blue) image is a three-

dimensional byte array that stores a unique color value for

each pixel. RGB image arrays consist of width, height, and

three-color channels. An RGB image can be regarded

logically as three independent images (a red scale image, a

green scale image, and a blue scale image) placed on top

each other, as shown in Fig. 2.

Fig. 2. RGB Image Representation.

An image in RGB format increases the complexity of the

model. This is why it is preferable to use grayscale images

over RGB to simplify computation.

A grayscale image, as illustrated in Fig. 3, it does not contain

color information rather, it contains only information related

assessment of BMD in the study women was also used to

decide whether densitometry testing was required (based on a

questionnaire with questions largely regarding dietary habits).

Regression trees were used to identify which elements were

most crucial, and SVMs were used to build a mathematical

model that reflected the relationship. The most important

things for postmenopausal women to do to prevent bone

density loss include consuming extra calcium, getting enough

sun, managing their weight, exercising regularly, and eating

enough calories [14].

The authors in [15], based on identified risk factors,

established a modern, effective bone disease prediction

model. Then, using Pre-training and fine-tuning, it was

possible to identify the early risk factors for determining the

start of bone problems. During the pre-training phase, the most

important risk factors are combined with model parameters to

calculate contrastive divergence, which reduces record size.

The results of the previous phase were compared using the

ground truth values "g1" and "g2," where g1 represented

osteoporosis and g2 represented a rate of bone loss. Deep

Belief Network (DBN) was used to generate the model, which

was then compared to models created before and after critical

feature identification. The study's findings suggested that

including relevant variables could increase the

prediction model's effectiveness, [15].

The authors in [16] created and assessed DL approaches

for osteoporosis classification using Dental Panoramic

Radiographs (DPR). In this work, various CNN models for

osteoporosis discriminating accuracy were tested using

panoramic radiograph pictures that had been categorized

based on BMD value (T-score). The effects of transfer

learning and fine-tuning a deep CNN model were also

evaluated in terms of classification performance. Deep CNN’s

have been found to be useful for classifying images, but

because they need a lot of training data, it is challenging to

apply them to radiographic medical imaging data. Transfer

learning is a popular strategy for training deep CNN’s without

"overfitting" when the target dataset is significantly smaller

than the basis dataset [17].

3. Materials and Methods

3.1 Data Acquisition

The dataset was gotten from a public dataset repository for

machine learning called Kaggle. The name of the Kaggle repo

is “Osteoporosis Knee X-ray Dataset”, version 1, uploaded on

the 16th of September, 2021 accessible via

www.kaggle.com/stevepython/datasets. The number

of images was increased using data augmentation. Fig. 1

shows two images from the dataset indicating osteoporosis

cases and normal cases.

3.2 Image Scaling

3.3 Image Formats

WSEAS TRANSACTIONS on ELECTRONICS

DOI: 10.37394/232017.2022.13.7

Usman Bello Abubakar, Moussa Mahamat Boukar,

Steve Adeshina, Senol Dane

E-ISSN: 2415-1513

Volume 13, 2022

to pixel brightness. The grayscale data matrix values are then

used to indicate intensities.

Fig. 3 Grayscale Image Representation

Overfitting can be reduced by using a technique called

data augmentation. “Overfitting occurs when a“model learns

a function with a relatively large variance to perfectly model

the training data” [21]. For this study, the Keras

ImageDataGenerator python class was used to perform data

augmentation using a variety of augmentation techniques as

itemized below:

1. Standardization

2. Rotation

3. Shifts

4. Brightness changes, among others

Fig. 4. GoogLeNet Model [24].

Fig. 5. VGG-16 Model [20].

3.4 Data Augmentation

The Keras ImageDataGenerator class is intended to give

real-time data augmentation, which is said to be its key

advantage. Every epoch, the model is given fresh versions of

the images thanks to the ImageDataGenerator class.

3.5 Transfer Learning Model Architecture

Three transfer learning model architectures were applied:

GoogLeNet, VGG-16, and ResNet-50. All the layers of the

pre-trained model were made to be non-trainable. However,

some of the layers could be re-trained to increase performance

but at the cost of a higher chance of model overfitting. For this

model, as the loss metric, binary_crossentropy was used as the

dataset target has two classes (i.e., binary classification

problem). RMSprop is the chosen optimizer, and its learning

rate is 0.001. Each model underwent 10 epochs of training.

The GoogleNet is a 22-layer deep convolutional

neural architecture that addressed computer vision issues such

as object recognition and image classification in the

ImageNet It has achieved 93.9% accuracy in the top 5

results [18], [19]. Fig. 4 shows the GoogLeNet model's

architecture.

VGG16 is CNN architecture. Having 16 layers and

is distinguished by its simplicity by having just a stack

of 33 convolutional layers on top of each other, with

the max-pooling layers handling the rising depth and

volume size. A softmax layer comes after two fully linked

layers with 4096 nodes each [20]. In ImageNet, the

VGG16 model obtained top-5 test accuracy “of 92.7%.

Fig. 5. Shows the VGG-16” model architecture [20].

WSEAS TRANSACTIONS on ELECTRONICS

DOI: 10.37394/232017.2022.13.7

Usman Bello Abubakar, Moussa Mahamat Boukar,

Steve Adeshina, Senol Dane

E-ISSN: 2415-1513

Volume 13, 2022

We experimented with the osteoporosis patient knee x-

rays dataset. In all transfer learning models, the dataset was

split into 80:20 ratio for training and testing. The overall

accuracy obtained for all the classifiers on the dataset is

summarized in Table II. Each model underwent 10 epochs of

training. For all models, as the loss metric,

binary_crossentropy was used as the dataset target has two

classes (i.e., binary classification problem). RMSprop is the

chosen optimizer, and its learning rate is 0. 001. The Keras

evaluate function was invoked on the compiled model with the

test data as an argument to evaluate the accuracy of the

models.

The confusion matrix for the transfer learning models is

presented in “Fig. 6, Fig. 7, Fig. 8, and Fig. 9”

Fig. 6. Confusion Matrix for GoogLeNet Model on Grayscale.

Fig. 7. Confusion Matrix for VGG-16 on Grayscale

Fig. 8. Confusion Matrix for GoogLeNet on RGB

Fig. 9. Confusion Matrix for VGG-16 on RGB

The following deep learning classification metric were

used to further understand the performance of the models on

the two modalities of the image formats.

Sensitivity = 𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒

𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒+𝑓𝑎𝑙𝑠𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 (1)

Specificity = 𝑡𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒

𝑡𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 +𝑓𝑎𝑙𝑠𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 (2)

Accuracy” = 𝑡𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒+𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒

𝑎𝑙𝑙 𝑐𝑎𝑠𝑒𝑠 (3)

Precision” = 𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒

𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒+𝑓𝑎𝑙𝑠𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒 (4)

F1” = 2 x 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 𝑥 𝑟𝑒𝑐𝑎𝑙𝑙

𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑟𝑒𝑐𝑎𝑙𝑙 (5)

Table II and Table III show the accuracy, epoch, and

training time of the models on grayscale images and RGB

4. Results

4.1 Confusion Matrix

4.2 Classification Metrics

WSEAS TRANSACTIONS on ELECTRONICS

DOI: 10.37394/232017.2022.13.7

Usman Bello Abubakar, Moussa Mahamat Boukar,

Steve Adeshina, Senol Dane

E-ISSN: 2415-1513

Volume 13, 2022

images respectively. Table IV provides a comparison of our

work with similar works on accuracy

TABLE II. RESULTS OBTAINED FOR GRAYSCALE IMAGES

Se/Re

F1-

Score

Time

(minutes)

GoogLeNet

0.90

0.91

0.90

0.89

0.90

VGG-16

0.87

0.86

0.87

0.86

TABLE III. RESULTS OBTAINED FOR RGB IMAGES

Se/Re

F1-

Score

Time

(minutes)

GoogLeNet

0.84

0.85

0.84

0.83

0.85

VGG-16

0.79

0.81

0.78

0.79

*AC: ACCURACY, SE: SENSITIVITY RE: RECALL, SP: SPECIFICITY

TABLE IV. COMPARISON WITH OTHER SIMILAR WORK

Paper

Method

Our Paper

GoogLeNet

0.90

0.91

0.90

Our Paper

VGG-16

0.87

0.86

N. Yamamoto et

al., [21]

ResNet-18

0.79

0.86

N. Yamamoto et

al., [21]

ResNet-34

0.84

0.88

0.86

K. S. Lee et al., [22]

VGG-16-

Fine-Tuning

0.84

0.90

0.81

K. S. Lee et al., [22]

CNN with 3

layers

0.66

0.68

0.65

S. Sukegawa et al., [23]

ResNet-50

0.83

0.75

0.90

*AC: ACCURACY, SE: SENSITIVITY, SP: SPECIFICITY

The behavior of the models in terms of speed over several

iterations can be better visualized in the following figures.

Fig. 7. Time Chart for Grayscale Training.

Fig. 8. Time Chart for Grayscale Training.

Fig. 9, Fig. 10, Fig. 11, and Fig. 12 shows the training and

validation accuracy for all deep learning models used.

Fig. 9. Train/Accuracy Graph for GoogLeNet on Grayscale Images.

Fig. 10. Train/Accuracy Graph for VGG-16 on Grayscale Images.

GoogLeNet VGG-16

Training Time Chart

GoogLeNet VGG-16

Training Time Chart

4.3 Training Time Chart

4.4 Training and Validation Accuracy Graph

WSEAS TRANSACTIONS on ELECTRONICS

DOI: 10.37394/232017.2022.13.7

Usman Bello Abubakar, Moussa Mahamat Boukar,

Steve Adeshina, Senol Dane

E-ISSN: 2415-1513

Volume 13, 2022

Fig. 11. Train/Accuracy Graph for GoogLeNet on RGB Images.

Fig. 12. Train/Accuracy Graph for VGG-16 on RGB Images.

Five folds were randomly selected from the training

dataset of the chosen images. This prevented bias or

overfitting while performing a 5-fold cross-validation on the

model training. The dataset was split into independent

training and validation sets within each fold using an 80 to 20

split. A validation set that was completely different from the

other training folds was chosen in order to assess the training

state throughout training. Once one model training phase was

complete, the other independent fold was utilized as a

validation set, and the previous validation set was recycled as

part of the training set to evaluate the model training.

This study uses a dataset of knee radiographs (i.e., knee-

Xray images) to apply and compare the training time of two

robust transfer learning model algorithms. The transfer

learning models applied were GoogLeNet, and VGG-16.

ImageDataGenerator was used to augment the dataset and

increase the number of training data to provide a variety of

images for the models. The dataset was split into two

subcategories using python opencv library: Grayscale Images

and Red Green Blue (RGB) images. From the scikit learn

python analysis, the training time of the GoogLeNet model on

grayscale images and RGB images was 42minutes and 50

minutes respectively. The VGG-16 model training time on

grayscale images and RGB images was 37 minutes and 44

minutes respectively. In addition, to compare the diagnostic

performance of the two models, several state-of-the-art neural

networks metric was used.

Osteoporosis is caused not just by low bone mineral

density, but also by other factors such as age, gender, weight,

height, and so on. These are clinically important risk factors

for osteoporosis. For future work, we would like to extend our

methods by adding patient variables such as age, and gender,

amongst others, as clinical covariates to create an ensemble

model with the transfer learning models.

[1] S. Gschmeissner and S. Photo Library, “Diagnosis,

assessment and management of osteoporosis,”

Prescriber, vol. 31, no. 1, pp. 14–19, Jan. 2020, doi:

10.1002/PSB.1815.

[2] U. Föger-Samwald, P. Dovjak, U. Azizi-Semrad, K.

Kerschan-Schindl, and P. Pietschmann,

“Osteoporosis: Pathophysiology and therapeutic

options,” EXCLI Journal, vol. 19, p. 1017, 2020,

doi: 10.17179/EXCLI2020-2591.

[3] E. M. Curtis, R. J. Moon, N. C. Harvey, and C.

Cooper, “The impact of fragility fracture and

approaches to osteoporosis risk assessment

worldwide,” Bone, vol. 104, pp. 29–38, Nov. 2017,

doi: 10.1016/j.bone.2017.01.024.

[4] J. R. Curtis et al., “Longitudinal Trends in Use of

Bone Mass Measurement Among Older Americans,

1999–2005,” Journal of Bone and Mineral

Research, vol. 23, no. 7, pp. 1061–1067, Jul. 2008,

doi: 10.1359/JBMR.080232.

[5] H. K. Lim, H. il Ha, S. Y. Park, and K. Lee,

“Comparison of the diagnostic performance of CT

Hounsfield unit histogram analysis and dual-energy

X-ray absorptiometry in predicting osteoporosis of

the femur,” European Radiology 2018 29:4, vol. 29,

no. 4, pp. 1831–1840, Sep. 2018, doi:

10.1007/S00330-018-5728-0.

[6] S. Jang, P. M. Graffy, T. J. Ziemlewicz, S. J. Lee, R.

M. Summers, and P. J. Pickhardt, “Opportunistic

osteoporosis screening at routine abdominal and

Thoracic CT: Normative L1 trabecular attenuation

values in more than 20 000 adults,” Radiology, vol.

291, no. 2, pp. 360–367, May 2019, doi:

10.1148/RADIOL.2019181648/ASSET/IMAGES/L

ARGE/RADIOL.2019181648.TBL1.JPEG.

[7] S. Lee, E. K. Choe, H. Y. Kang, J. W. Yoon, and H.

S. Kim, “The exploration of feature extraction and

machine learning for predicting bone density from

simple spine X-ray images in a Korean population,”

Skeletal Radiology, vol. 49, no. 4, pp. 613–618, Apr.

2020, doi: 10.1007/S00256-019-03342-6.

[8] U. Ferizi, S. Honig, and G. Chang, “Artificial

intelligence, osteoporosis and fragility fractures,”

Curr Opin Rheumatol, vol. 31, no. 4, pp. 368–375,

Jul. 2019, doi: 10.1097/BOR.0000000000000607.

[9] J. J. Hwang et al., “Strut analysis for osteoporosis

detection model using dental panoramic

radiography,” Dentomaxillofacial Radiology, vol.

46, no. 7, 2017, doi: 10.1259/DMFR.20170006.

4.5 Cross Validation

5. Conclusion

References

WSEAS TRANSACTIONS on ELECTRONICS

DOI: 10.37394/232017.2022.13.7

Usman Bello Abubakar, Moussa Mahamat Boukar,

Steve Adeshina, Senol Dane

E-ISSN: 2415-1513

Volume 13, 2022

[10] K. S. Lee, S. K. Jung, J. J. Ryu, S. W. Shin, and J.

Choi, “Evaluation of Transfer Learning with Deep

Convolutional Neural Networks for Screening

Osteoporosis in Dental Panoramic Radiographs,”

undefined, vol. 9, no. 2, Feb. 2020, doi:

10.3390/JCM9020392.

[11] H. P. Dimai et al., “Assessing the effects of long-

term osteoporosis treatment by using conventional

spine radiographs: results from a pilot study in a

sub-cohort of a large randomized controlled trial,”

Skeletal Radiology, vol. 48, no. 7, pp. 1023–1032,

Jul. 2019, doi: 10.1007/S00256-018-3118-Y.

[12] A. S. Areeckal, N. Jayasheelan, J. Kamath, S.

Zawadynski, M. Kocher, and S. David S, “Early

diagnosis of osteoporosis using radiogrammetry and

texture analysis from hand and wrist radiographs in

Indian population,” Osteoporosis International, vol.

29, no. 3, pp. 665–673, Mar. 2018, doi:

10.1007/S00198-017-4328-1.

[13] N. Tecle, J. Teitel, M. R. Morris, N. Sani, D. Mitten,

and W. C. Hammert, “Convolutional Neural

Network for Second Metacarpal Radiographic

Osteoporosis Screening,” The Journal of Hand

Surgery, vol. 45, no. 3, pp. 175–181, Mar. 2020,

doi: 10.1016/J.JHSA.2019.11.019.

[14] C. Ordóñez, J. M. Matías, J. F. de Cos Juez, and P.

J. García, “Machine learning techniques applied to

the determination of osteoporosis incidence in post-

menopausal women,” Mathematical and Computer

Modelling, vol. 50, no. 5–6, pp. 673–679, Sep.

2009, doi: 10.1016/J.MCM.2008.12.024.

[15] M. Saranya, M. Sc, M. Phil, and K. Sarojini, “An

Improved and Optimal Prediction of Bone Disease

Based On Risk Factors.” [Online]. Available:

www.ijcsit.com

[16] K. S. Lee, S. K. Jung, J. J. Ryu, S. W. Shin, and J.

Choi, “Evaluation of Transfer Learning with Deep

Convolutional Neural Networks for Screening

Osteoporosis in Dental Panoramic Radiographs,”

undefined, vol. 9, no. 2, Feb. 2020, doi: 10.3390/

JCM9020392.

[17]

[18]

[19]

[20]

[21]

[22]

[23] S. Sukegawa et al., “Identification of osteoporosis

using ensemble deep learning model with panoramic

radiographs and clinical covariates,” Scientific

Reports 2022 12:1, vol. 12, no. 1, pp. 1–10, Apr.

2022, doi: 10.1038/s41598-022-10150-x.

[24] “Inception V3 Model Architecture.”

https://iq.opengenus.org/inception-v3-model-

architecture/ (accessed Mar. 30, 2022).

WSEAS TRANSACTIONS on ELECTRONICS

DOI: 10.37394/232017.2022.13.7

Usman Bello Abubakar, Moussa Mahamat Boukar,

Steve Adeshina, Senol Dane

E-ISSN: 2415-1513

Volume 13, 2022

[17] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson,

“How transferable are features in deep neural

networks?”.

[18] C. Shorten and T. M. Khoshgoftaar, “A survey on

Image Data Augmentation for Deep Learning,”

Journal of Big Data, vol. 6, no. 1, Dec. 2019, doi:

10.1186/S40537-019-0197-0.

[19] “Transfer Learning using Inception-v3 for Image

Classification | by Tejan Irla | Analytics Vidhya |

Medium.” https://medium.com/analytics-

vidhya/transfer-learning-using-inception-v3-for-

image-classification-86700411251b (accessed Apr.

04, 2022).

[20] “What is VGG16? — Introduction to VGG16 | by

Great Learning | Medium.”

https://medium.com/@mygreatlearning/what-is-

vgg16-introduction-to-vgg16-f2d63849f615

(accessed Mar. 30, 2022).

[21] N. Yamamoto et al., “Deep learning for osteoporosis

classification using hip radiographs and patient

clinical covariates,” Biomolecules, vol. 10, no. 11,

pp. 1–13, Nov. 2020, doi:

10.3390/BIOM10111534.

[22] K. S. Lee, S. K. Jung, J. J. Ryu, S. W. Shin, and J.

Choi, “Evaluation of Transfer Learning with Deep

Convolutional Neural Networks for Screening

Osteoporosis in Dental Panoramic Radiographs,”

Journal of Clinical Medicine, vol. 9, no. 2, Feb.

2020, doi: 10.3390/JCM9020392.

Conflicts of Interest

The author(s) declare no potential conflicts of

interest concerning the research, authorship, or

publication of this article.

Contribution of individual authors to

the creation of a scientific article

(ghostwriting policy)

The author(s) contributed in the present

research, at all stages from the formulation

of the problem to the final findings

and solution.

Sources of funding for research

presented in a scientific article or

scientific article itself

No funding was received for conducting this

study.

Creative Commons

Attribution License

4.0 (Attribution 4.0 Int

ernational, CC

BY 4.0))

This article is published under the terms of the

Creative Commons Attribution License 4.0

https://creativecommons.org/licenses/by/4.0/deed.en_US