Osteoporosis Detection Using Deep Learning
0 likes169 views
The document discusses the use of a convolutional neural network (CNN) model for the early diagnosis of osteoporosis through analysis of bone radiography images. It highlights the challenges in distinguishing healthy from osteoporotic subjects due to similarities in image textures and emphasizes the necessity of accurate automated diagnosis to address the global health burden posed by osteoporosis-related fractures. The model achieves a classification accuracy of 79.3% using a transfer learning approach with a pre-trained CNN on a limited dataset.
![18 International Journal for Modern Trends in Science and Technology
Sylvia Grace J, Dinesh Kumar S, Gautam R and Mark Sachin K : Osteoporosis Detection Using Deep Learning
and fastest way for the doctors to study the injuries
of joints and bones.
Texture Characterization of Bone radiograph
images is a challenge in the osteoporosis diagnosis.
Osteoporosis can be defined as a skeletal disorder
characterized by a compromise on bone strength
predisposing to an increase in the occurrence of
fracture. The most common method for
osteoporosis diagnosis is to estimate Bone Mineral
Density (BMD) by dual-energy X-ray
absorptiometry. However, BMD alone represents
only 60% of fracture prediction. The
characterization of trabecular bone
microarchitecture has been recognized as an
important factor and completes the osteoporosis
diagnosis using BMD, but it cannot be routinely
obtained by noninvasive methods and requires a
bone biopsy with histomorphometric analysis. 2-D
texture analysis offers an easy way to evaluate
bone structure on conventional radiography
images. The evaluation of osteoporotic disease from
bone radiograph images presents a major
challenge for pattern recognition and medical
applications. Textured images from the bone
micro-architecture of healthy and osteoporotic
subjects show a high degree of similarity, thus
drastically increasing the difficulty of classifying
such textures. Fig. 1 & 2 shows the bone texture
similarities of control and osteoporotic images.
Fig. 1: Osteoporotic Patient X Ray Image
Fig. 2: Control Patient X Ray Image
II. RELATED WORKS
A variety of these techniques, including Artificial
neural networks (ANNs), Bayesian Networks (BNs),
SupportVector Machines (SVMs) and Decision
Trees (DTs) have been widely implemented in
medical research for the design and development of
predictive models, resulting in accurate and
effective decision making. Even though it is evident
that the use of Machine Learning methods can
improve our understanding of Osteoporosis
detection, an appropriate level of validation is
needed in order for these methods to be considered
in everyday clinical practice.
Michel Kocher [1] has discussed recent and
state-of-the-art advances in imaging techniques for
the diagnosis of osteoporosis and fracture risk
assessment. The author has also talked about
segmentation methods used to segment the region
of interest and texture analysis methods used for
classification of osteoporotic and healthy subjects.
Also, challenges posed by the current diagnostic
tools have been studied and feasible solutions to
circumvent the limitations are discussed.
Costin Florian Ciu܈del [2] has discussed
physics-based models, employing finite element
analysis (FEA) which have shown great promise in
being able to non-invasively estimate
biomechanical quantities of interest in the context
of osteoporosis. The limitation is that these models
have high computational demand which limits
their clinical adoption. The paper discusses a deep
learning model based on a convolutional neural
network (CNN) for predicting average strain as an
alternative to physics-based approaches. The
model is trained on a large database of
synthetically generated cancellous bone anatomies,
where the target values are computed using the
physics-based FEA model. The performance of the
trained model was assessed by comparing the
predictions against physics-based computations
on a separate test data set. Correlation between
deep learning and physics-based predictions was
very good (0.895, p < 0.001), and no systematic
bias were found in BlandAltman analysis. The CNN
model also performed better than the previously
introduced Support Vector Machine (SVM) model
which relied on handcrafted features (correlation
0.847, p < 0.001). Compared to the physics-based
computation, average execution time was reduced
by more than 1000 times, leading to a real-time
assessment of average strain.
Kazuhiro Hatano [3] has discussed visual
screening using Computed Radiography (CR)
images is an effective method for osteoporosis,
however, there are many similar diseases that
exhibit a state of low bone mass. In this paper, we
propose an automatic identification method of
osteoporosis from phalanges CR images. In the
proposed method, we implement a classifier based
on Deep Convolutional Neural Network (DCNN),](https://image.slidesharecdn.com/4-190331165925/85/Osteoporosis-Detection-Using-Deep-Learning-2-320.jpg)
![19 International Journal for Modern Trends in Science and Technology
Sylvia Grace J, Dinesh Kumar S, Gautam R and Mark Sachin K : Osteoporosis Detection Using Deep Learning
and identify unknown CR images as normal or
abnormal. For training and evaluating of CNN, we
use pseudocolor images.
MS Kavitha [5] has discussed the diagnostic
efficacies achieved by the application of an RBF
kernel-SVM showed that CAD system was accurate
and effective for identifying women with low BMD.
The author claims the SVM method can be a
reliable choice for the proposed system because it
is fast and specific for classification, using dental
panoramic radiographs, of postmenopausal
women with low BMD. Based on highly satisfactory
sensitivity and specificity results, the proposed
system is expected to be a helpful tool for
classifying women with low BMD and is also
expected to provide a second diagnosis that may
reduce misdiagnoses.
Peng Chu [6] has proposed an image-based
osteoporosis classification method using dental
panoramic radiography dataset. The method
combines multiple ROI information by using a
two-stage classification model, the first stage is a
linear SVM model which is constructed with
respect to a grouped region and given feature. The
probability outputs of the first stage SVM models
and the new SVM model are combined together.
The experimental results show that the proposed
method with the HOG feature achieves an accuracy
of 72.5% with a low p-value (0.0164). The author
concludes that the combination of image feature
analysis and machine learning techniques on
panoramic radiography images has the potential
for osteoporosis detection.
III. METHODS AND MATERIALS
In Machine Learning, classification and object
detection using Convolutional Neural Networks
(CNNs) have shown better performance. Usage of
Artificial Neural Networks (ANNs), which are
inspired by the human brain, can also be used for
classification and prediction. An ANN mainly has
three input layers, an output and a hidden layer
with activation function on hidden and output
layers. These layers consist of a number of
interconnected nodes, the number of weights in
between the nodes will increase rapidly, which is
an issue with the ANN. CNN a translational
invariant of the neural networks which consists of
some convolutional and fully connected layers.
CNN uses several small filters on the input and
then subsampling the space of filter activations
until there is a good number of high-level features.
A Convolutional Neural Network is biologically
influenced variants of multi-layer feed forward
model that is currently widely using in image
classification and recognition tasks. The layers of
CNN are mainly classified into four layers or
operations. The first layer is a convolutional layer,
which is based on a mathematical approach called
convolution, which takes a kernel and applied it to
all over an input image to generate a filtered image.
The second operation is Pooling or sub-sampling
which reduces the dimensionality of the feature
map although keep the most important
information for the next convolution layer. The
third layer is an activation function layer, which
applies activation function elementwise and got the
output by checking whether the neurons are fired
or not. ReLU, sigmoid functions, tanh functions
etc., are some of the common activations functions.
The last layer is a fully connected layer, where all
the neurons from the previous layer are connected
to all nodes in the adjacent layers which is the
same as the fully connected multilayer perceptron.
CNN must learn lots of weights and thus, we need a
huge amount of data for training, but in the
medical domain, we don’t get enough data to be
used for training from scratch. To counter this
problem for the training of CNN an alternative
approach Transfer learning can be used. Transfer
learning is an approach where knowledge leaned
from the previous task can be applied to some new
task domain. Here, we have used transfer learning
concept, by using pre-trained CNN which is trained
on ImageNet dataset. ImageNet is a popular
database for images which consists of more than
13 million images of 1000 distinct object classes.
We have used a pre-trained CNN architecture
(VGG16) as described in as one of Chatfield’s [8]](https://image.slidesharecdn.com/4-190331165925/85/Osteoporosis-Detection-Using-Deep-Learning-3-320.jpg)
![20 International Journal for Modern Trends in Science and Technology
Sylvia Grace J, Dinesh Kumar S, Gautam R and Mark Sachin K : Osteoporosis Detection Using Deep Learning
work. The CNN has five convolution layer and
followed by three fully connected layers. We are
using a pre-trained network to extract features
from the last hidden layer after applying the
activation function (post relu) [9]. The X-Ray
images used in the experiment are different from
the images in the ImageNet database, but we are
hypothesizing some useful texture features might
exist. The input image size is 224x224 for the CNN
architectures so we use bicubic interpolation to
resize the input images. The x-ray images are
grayscale, so we modified the code and extracting
features using only the R channel and ignoring B
and G channel. The deep features that we are
extracting are of 4096 dimensions.
IV.EXPERIMENT
The dataset used in this experiment consists of
174 bone texture X-Ray images obtained from
IEEE-ISBI 2014 competition. In the given dataset,
58 images are using for testing and don’t have any
labels or class given and 116 images are equally
subdivided into osteoporotic and control cases.
We have used a pre-trained CNN model in our
experiment. The model required 224x224 input
image, so using bi-cubic interpolation method we
have resized the images. After resizing, the size of
extracted deep features vector from each X-Ray
images were 4096 and then classifying it using the
features. The best result of 79.3% was obtained
using post-relu features from the VGG.
V. CONCLUSION
Classifying Osteoporosis by just considering the
x-ray images is very difficult as the images
obtained from the osteoporotic patient looks very
similar to that of the healthy patient. We have
proposed CNN for classification and extracting
features. However, training a CNN requires a huge
amount of data. But we had access to only a very
small number of training data, which is insufficient
to train a CNN. To solve this problem, we have used
a transfer learning approach – pre-trained CNN
model. The pre-trained CNNs are trained on
ImageNet data and we are using it to extract
features from the last hidden layer after applying
the activation function. On the training set, the
best result of 79.3% was achieved. The future
scope of the project is to work more on deep feature
to generate a better classification result on blind
test data and tune the CNN to give better results.
REFERENCES
[1] Anu Shaju Areeckal, Student Member, IEEE, Michel
Kocher, Sumam David S., Senior Member, IEEE
(2018), “Current and Emerging Diagnostic
Imaging-Based Techniques for Assessment of
Osteoporosis and Fracture Risk” in IEEE
[2] Costin Florian Ciu܈del, Anamaria Vizitiu, Florin
Moldoveanu, Constantin Suciu and Lucian Mihai
Itu (2017), “Towards deep learning based estimation
of fracture risk in osteoporosis patients” in 40th
International Conference on Telecommunications
and Signal Processing (TSP)
[3] Kazuhiro Hatano, Seiichi Murakami, Huimin Lu,
Joo Kooi Tan, Hyoungseop Kim and Takatoshi Aoki
(2017), “Classification of Osteoporosis from
phalanges CR images based on DCNN” in 17th
International Conference on Control, Automation
and Systems (ICCAS).
[4] C. F. Ciuşdel, A. Vizitiu, F. Moldoveanu, C. Suciu
and L. M. Itu, "Towards deep learning based
estimation of fracture risk in osteoporosis patients,"
2017 40th International Conference on
Telecommunications and Signal Processing (TSP),
Barcelona, 2017
[5] Kavitha MS, Asano A, Taguchi A, Kurita T, Sanada
M, “Diagnosis of osteoporosis from dental
panoramic radiographs using the support vector
machine method in a computer-aided system”, BMC
Med Imaging, 2012
[6] Peng Chu, Chunjuan Bo, Xin Liang, Jie Yang,
Vasileios Megalooikonomou, Fan Yang, Bingyao
Huang, Xinyi Li, Haibin Ling, "Using Octuplet
Siamese Network For Osteoporosis Analysis On
Dental Panoramic Radiographs", IEEE Engineering
in Medicine and Biology Society (EMBC) 2018 40th
Annual International Conference of the, pp.
2579-2582, 2018
[7] Yu L, Liu H. Feature selection for high-dimensional
data: A fast correlation-based filter solution.
International Conference on Machine Learning.
2003;3:856-863.
[8] K. Chatfield, K. Simonyan, A. Vedaldi and A.
Zisserman,” Return of the devil in the details:
Delving deep into convolutional nets.”, 2014
[9] R. Paul, S. Hawkins, Y. Balagurunathan, M.
Schabath, R. Gillies, L. Hall, D. Goldgof, ”Deep
Feature Transfer Learning in Combination with
Traditional Features Predicts Survival among
Patients with Lung Adenocarcinoma”, Tomography
Journal, Special QIN Issue, 2016](https://image.slidesharecdn.com/4-190331165925/85/Osteoporosis-Detection-Using-Deep-Learning-4-320.jpg)
Osteoporosis Detection Using Deep Learning
- 1. 17 International Journal for Modern Trends in Science and Technology Osteoporosis Detection Using Deep Learning Sylvia Grace J1 | Dinesh Kumar S2 | Gautam R3 | Mark Sachin K4 1Assistant Professor, Department of Computer Science & Engineering, KCG College of Technology, Chennai 2 3 4 Final Year, Department of Computer Science & Engineering, KCG College of Technology, Chennai To Cite this Article Sylvia Grace J, Dinesh Kumar S, Gautam R and Mark Sachin K, “Osteoporosis Detection Using Deep Learning”, International Journal for Modern Trends in Science and Technology, Vol. 05, Issue 03, March 2019, pp.-17-20. Article Info Received on 21-Feb-2019, Revised on 17-March-2019, Accepted on 28-March-2019. Osteoporosis is a bone disorder which occurs due to low bone mass, degradation of bone micro-architecture and high susceptibility to fracture. It is a major health concern across the world, especially in elderly people. Osteoporosis can cause spinal or hip fractures that may lead to socio-economic burden and high morbidity. Therefore, there is a need for the early diagnosis of osteoporosis and predicting the presence of the fracture. We introduce a Convolutional Neural Network model to effectively diagnose osteoporosis in bone radiography data. Automated diagnosis from digital radiographs is very challenging since the scans of healthy and osteoporotic subjects show little or no visual differences. In this paper, we have proposed a model to separate healthy from osteoporotic subjects using high dimensional textural feature representations computed from radiography images. CNN can help us bring the use of structural MRI measurements of bone quality into clinical practice for the detection of Osteoporosis as it gives high accuracy. KEYWORDS: CNN, MRI, CT Scan, Osteoporosis Copyright © 2019 International Journal for Modern Trends in Science and Technology All rights reserved. I. INTRODUCTION Osteoporosis is a bone condition caused by a reduction in bone mass and degeneration of bone structure, that leads to high susceptibility to fragility fractures. Osteoporosis-related fracture is a major global health risk, affecting one in three women and one in five men over the age of 50. According to the Asia-Pacific Regional Audit in 2013, osteoporosis accounts for more hospitalization than diabetes, myocardial infarction, and breast cancer, in women above the age of 45 years. It is more prevalent among the elderly population, especially postmenopausal women. With a rise in the aging population, there will be a substantial rise in the incidence of fractures. It is projected that by 2050, at least one-third of the world population will be aged over 50 years and in Asia alone, a 7.6-fold increase in aging population is expected, that may result in more than 50% of the global fractures to occur in Asia. Bones are the rigid organs in the human body which protect important organs such as brain, heart, lungs and other internal organs. The human body has 206 bones with various shapes, size, and structures. There are varied types of medical imaging tools that are available to detect different types of abnormalities such as Computed Tomography (CT), X-ray and Magnetic Resonance Imaging (MRI). CT and X-rays are most frequently used for fracture diagnosis because it is the easiest ABSTRACT Available online at: http://www.ijmtst.com/vol5issue03.html International Journal for Modern Trends in Science and Technology ISSN: 2455-3778 :: Volume: 05, Issue No: 03, March 2019
- 2. 18 International Journal for Modern Trends in Science and Technology Sylvia Grace J, Dinesh Kumar S, Gautam R and Mark Sachin K : Osteoporosis Detection Using Deep Learning and fastest way for the doctors to study the injuries of joints and bones. Texture Characterization of Bone radiograph images is a challenge in the osteoporosis diagnosis. Osteoporosis can be defined as a skeletal disorder characterized by a compromise on bone strength predisposing to an increase in the occurrence of fracture. The most common method for osteoporosis diagnosis is to estimate Bone Mineral Density (BMD) by dual-energy X-ray absorptiometry. However, BMD alone represents only 60% of fracture prediction. The characterization of trabecular bone microarchitecture has been recognized as an important factor and completes the osteoporosis diagnosis using BMD, but it cannot be routinely obtained by noninvasive methods and requires a bone biopsy with histomorphometric analysis. 2-D texture analysis offers an easy way to evaluate bone structure on conventional radiography images. The evaluation of osteoporotic disease from bone radiograph images presents a major challenge for pattern recognition and medical applications. Textured images from the bone micro-architecture of healthy and osteoporotic subjects show a high degree of similarity, thus drastically increasing the difficulty of classifying such textures. Fig. 1 & 2 shows the bone texture similarities of control and osteoporotic images. Fig. 1: Osteoporotic Patient X Ray Image Fig. 2: Control Patient X Ray Image II. RELATED WORKS A variety of these techniques, including Artificial neural networks (ANNs), Bayesian Networks (BNs), SupportVector Machines (SVMs) and Decision Trees (DTs) have been widely implemented in medical research for the design and development of predictive models, resulting in accurate and effective decision making. Even though it is evident that the use of Machine Learning methods can improve our understanding of Osteoporosis detection, an appropriate level of validation is needed in order for these methods to be considered in everyday clinical practice. Michel Kocher [1] has discussed recent and state-of-the-art advances in imaging techniques for the diagnosis of osteoporosis and fracture risk assessment. The author has also talked about segmentation methods used to segment the region of interest and texture analysis methods used for classification of osteoporotic and healthy subjects. Also, challenges posed by the current diagnostic tools have been studied and feasible solutions to circumvent the limitations are discussed. Costin Florian Ciu܈del [2] has discussed physics-based models, employing finite element analysis (FEA) which have shown great promise in being able to non-invasively estimate biomechanical quantities of interest in the context of osteoporosis. The limitation is that these models have high computational demand which limits their clinical adoption. The paper discusses a deep learning model based on a convolutional neural network (CNN) for predicting average strain as an alternative to physics-based approaches. The model is trained on a large database of synthetically generated cancellous bone anatomies, where the target values are computed using the physics-based FEA model. The performance of the trained model was assessed by comparing the predictions against physics-based computations on a separate test data set. Correlation between deep learning and physics-based predictions was very good (0.895, p < 0.001), and no systematic bias were found in BlandAltman analysis. The CNN model also performed better than the previously introduced Support Vector Machine (SVM) model which relied on handcrafted features (correlation 0.847, p < 0.001). Compared to the physics-based computation, average execution time was reduced by more than 1000 times, leading to a real-time assessment of average strain. Kazuhiro Hatano [3] has discussed visual screening using Computed Radiography (CR) images is an effective method for osteoporosis, however, there are many similar diseases that exhibit a state of low bone mass. In this paper, we propose an automatic identification method of osteoporosis from phalanges CR images. In the proposed method, we implement a classifier based on Deep Convolutional Neural Network (DCNN),
- 3. 19 International Journal for Modern Trends in Science and Technology Sylvia Grace J, Dinesh Kumar S, Gautam R and Mark Sachin K : Osteoporosis Detection Using Deep Learning and identify unknown CR images as normal or abnormal. For training and evaluating of CNN, we use pseudocolor images. MS Kavitha [5] has discussed the diagnostic efficacies achieved by the application of an RBF kernel-SVM showed that CAD system was accurate and effective for identifying women with low BMD. The author claims the SVM method can be a reliable choice for the proposed system because it is fast and specific for classification, using dental panoramic radiographs, of postmenopausal women with low BMD. Based on highly satisfactory sensitivity and specificity results, the proposed system is expected to be a helpful tool for classifying women with low BMD and is also expected to provide a second diagnosis that may reduce misdiagnoses. Peng Chu [6] has proposed an image-based osteoporosis classification method using dental panoramic radiography dataset. The method combines multiple ROI information by using a two-stage classification model, the first stage is a linear SVM model which is constructed with respect to a grouped region and given feature. The probability outputs of the first stage SVM models and the new SVM model are combined together. The experimental results show that the proposed method with the HOG feature achieves an accuracy of 72.5% with a low p-value (0.0164). The author concludes that the combination of image feature analysis and machine learning techniques on panoramic radiography images has the potential for osteoporosis detection. III. METHODS AND MATERIALS In Machine Learning, classification and object detection using Convolutional Neural Networks (CNNs) have shown better performance. Usage of Artificial Neural Networks (ANNs), which are inspired by the human brain, can also be used for classification and prediction. An ANN mainly has three input layers, an output and a hidden layer with activation function on hidden and output layers. These layers consist of a number of interconnected nodes, the number of weights in between the nodes will increase rapidly, which is an issue with the ANN. CNN a translational invariant of the neural networks which consists of some convolutional and fully connected layers. CNN uses several small filters on the input and then subsampling the space of filter activations until there is a good number of high-level features. A Convolutional Neural Network is biologically influenced variants of multi-layer feed forward model that is currently widely using in image classification and recognition tasks. The layers of CNN are mainly classified into four layers or operations. The first layer is a convolutional layer, which is based on a mathematical approach called convolution, which takes a kernel and applied it to all over an input image to generate a filtered image. The second operation is Pooling or sub-sampling which reduces the dimensionality of the feature map although keep the most important information for the next convolution layer. The third layer is an activation function layer, which applies activation function elementwise and got the output by checking whether the neurons are fired or not. ReLU, sigmoid functions, tanh functions etc., are some of the common activations functions. The last layer is a fully connected layer, where all the neurons from the previous layer are connected to all nodes in the adjacent layers which is the same as the fully connected multilayer perceptron. CNN must learn lots of weights and thus, we need a huge amount of data for training, but in the medical domain, we don’t get enough data to be used for training from scratch. To counter this problem for the training of CNN an alternative approach Transfer learning can be used. Transfer learning is an approach where knowledge leaned from the previous task can be applied to some new task domain. Here, we have used transfer learning concept, by using pre-trained CNN which is trained on ImageNet dataset. ImageNet is a popular database for images which consists of more than 13 million images of 1000 distinct object classes. We have used a pre-trained CNN architecture (VGG16) as described in as one of Chatfield’s [8]
- 4. 20 International Journal for Modern Trends in Science and Technology Sylvia Grace J, Dinesh Kumar S, Gautam R and Mark Sachin K : Osteoporosis Detection Using Deep Learning work. The CNN has five convolution layer and followed by three fully connected layers. We are using a pre-trained network to extract features from the last hidden layer after applying the activation function (post relu) [9]. The X-Ray images used in the experiment are different from the images in the ImageNet database, but we are hypothesizing some useful texture features might exist. The input image size is 224x224 for the CNN architectures so we use bicubic interpolation to resize the input images. The x-ray images are grayscale, so we modified the code and extracting features using only the R channel and ignoring B and G channel. The deep features that we are extracting are of 4096 dimensions. IV.EXPERIMENT The dataset used in this experiment consists of 174 bone texture X-Ray images obtained from IEEE-ISBI 2014 competition. In the given dataset, 58 images are using for testing and don’t have any labels or class given and 116 images are equally subdivided into osteoporotic and control cases. We have used a pre-trained CNN model in our experiment. The model required 224x224 input image, so using bi-cubic interpolation method we have resized the images. After resizing, the size of extracted deep features vector from each X-Ray images were 4096 and then classifying it using the features. The best result of 79.3% was obtained using post-relu features from the VGG. V. CONCLUSION Classifying Osteoporosis by just considering the x-ray images is very difficult as the images obtained from the osteoporotic patient looks very similar to that of the healthy patient. We have proposed CNN for classification and extracting features. However, training a CNN requires a huge amount of data. But we had access to only a very small number of training data, which is insufficient to train a CNN. To solve this problem, we have used a transfer learning approach – pre-trained CNN model. The pre-trained CNNs are trained on ImageNet data and we are using it to extract features from the last hidden layer after applying the activation function. On the training set, the best result of 79.3% was achieved. The future scope of the project is to work more on deep feature to generate a better classification result on blind test data and tune the CNN to give better results. REFERENCES [1] Anu Shaju Areeckal, Student Member, IEEE, Michel Kocher, Sumam David S., Senior Member, IEEE (2018), “Current and Emerging Diagnostic Imaging-Based Techniques for Assessment of Osteoporosis and Fracture Risk” in IEEE [2] Costin Florian Ciu܈del, Anamaria Vizitiu, Florin Moldoveanu, Constantin Suciu and Lucian Mihai Itu (2017), “Towards deep learning based estimation of fracture risk in osteoporosis patients” in 40th International Conference on Telecommunications and Signal Processing (TSP) [3] Kazuhiro Hatano, Seiichi Murakami, Huimin Lu, Joo Kooi Tan, Hyoungseop Kim and Takatoshi Aoki (2017), “Classification of Osteoporosis from phalanges CR images based on DCNN” in 17th International Conference on Control, Automation and Systems (ICCAS). [4] C. F. Ciuşdel, A. Vizitiu, F. Moldoveanu, C. Suciu and L. M. Itu, "Towards deep learning based estimation of fracture risk in osteoporosis patients," 2017 40th International Conference on Telecommunications and Signal Processing (TSP), Barcelona, 2017 [5] Kavitha MS, Asano A, Taguchi A, Kurita T, Sanada M, “Diagnosis of osteoporosis from dental panoramic radiographs using the support vector machine method in a computer-aided system”, BMC Med Imaging, 2012 [6] Peng Chu, Chunjuan Bo, Xin Liang, Jie Yang, Vasileios Megalooikonomou, Fan Yang, Bingyao Huang, Xinyi Li, Haibin Ling, "Using Octuplet Siamese Network For Osteoporosis Analysis On Dental Panoramic Radiographs", IEEE Engineering in Medicine and Biology Society (EMBC) 2018 40th Annual International Conference of the, pp. 2579-2582, 2018 [7] Yu L, Liu H. Feature selection for high-dimensional data: A fast correlation-based filter solution. International Conference on Machine Learning. 2003;3:856-863. [8] K. Chatfield, K. Simonyan, A. Vedaldi and A. Zisserman,” Return of the devil in the details: Delving deep into convolutional nets.”, 2014 [9] R. Paul, S. Hawkins, Y. Balagurunathan, M. Schabath, R. Gillies, L. Hall, D. Goldgof, ”Deep Feature Transfer Learning in Combination with Traditional Features Predicts Survival among Patients with Lung Adenocarcinoma”, Tomography Journal, Special QIN Issue, 2016