Classification of the trabecular bone structure of osteoporotic patients using machine vision

Accepted Manuscript Classification of the trabecular bone structure of osteoporotic patients using machine vision Anushikha Singh, Malay Kishore Dutta, Rachid Jennane, Eric Lespessailles PII:

S0010-4825(17)30337-2

DOI:

10.1016/j.compbiomed.2017.10.011

Reference:

CBM 2804

To appear in:

Computers in Biology and Medicine

Received Date: 17 January 2017 Revised Date:

22 September 2017

Accepted Date: 11 October 2017

Please cite this article as: A. Singh, M.K. Dutta, R. Jennane, E. Lespessailles, Classification of the trabecular bone structure of osteoporotic patients using machine vision, Computers in Biology and Medicine (2017), doi: 10.1016/j.compbiomed.2017.10.011. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

RI PT

Classification of the Trabecular Bone Structure of Osteoporotic Patients using Machine Vision Anushikha Singh1, Malay Kishore Dutta1, Rachid Jennane2, Eric Lespessailles3 1

M AN U

SC

Department of Electronics & Communication Engineering Amity University, Noida, Uttar Pradesh, India 2 Univ. Orléans, I3MTO Laboratory, EA 4708, 45067 Orléans, France 3 Univ. Orléans, I3MTO Laboratory, EA 4708 and Hospital of Orleans, 45067 Orléans, France Emails: [email protected], [email protected] [email protected], [email protected]

AC C

EP

TE D

Abstract: Osteoporosis is a common bone disease which often leads to fractures. Clinically, the major challenge for the automatic diagnosis of osteoporosis is the complex architecture of bones. The clinical diagnosis of osteoporosis is conventionally done using Dual-energy X-ray Absorptiometry (DXA). This method has specific limitations, however, such as the large size of the instrument, a relatively high cost and limited availability. The method proposed here is based on the automatic processing of X-ray images. The bone X-ray image was statistically processed and strategically reformed to extract discriminatory statistical features of different orders. These features were used for machine learning for the classification of two populations composed of osteoporotic and healthy subjects. Four classifiers - support vector machine (SVM), k-nearest neighbors, Naïve Bayes and artificial neural network - were used to test the performance of the proposed method. Tests were performed on X-ray images of the calcaneus bone collected from the hospital of Orleans. The results are significant in terms of accuracy and time complexity. Experimental results indicate a classification rate of 98% using an SVM classifier which is encouraging for automatic osteoporosis diagnosis using bone X-ray images. The low time complexity of the proposed method makes it suitable for real time applications. Keywords: Osteoporosis, Bone X-ray Images, Image analysis, Feature Extraction, Supervised Classification.

1 Introduction Osteoporosis is a Greek term meaning “porous bone”. It is the most common disease of bones, and causes loss of bone density or insufficient bone formation, making the bones more susceptible to fractures [1]. Osteoporosis may be considered as a silent disease as no symptoms occur in the early stages of the disease. This silent disease

1

ACCEPTED MANUSCRIPT

TE D

M AN U

SC

RI PT

affects a large part of the population worldwide from a certain age. According to the estimation of the World Health Organization (WHO), osteoporosis affects more than 75 million people in the USA, Europe and Japan [2] and is the reason for more than 2.3 million fractures yearly in Europe and the USA alone. Currently, it is estimated that more than 200 million people all over the world are affected from this disease of bone [3]. It is a very common disease in India as well because of increasing longevity and a larger proportion of the population over the age of 50 years. It is estimated that more than 50 million people in India suffer from osteoporosis [3]. Bone in humans is living, growing tissue which is mainly composed of protein, calcium and collagen. The strength and flexibility of bone depend on the combination of calcium and collagen. Osteoporosis is a consequence of a steady loss of collagen and calcium, which leads to a gradual decrease in bone density and deterioration of bone microarchitecture [4]. Bones become weak, and fractures thus occur frequently, mainly in the wrist, hip, spine or femur. Fig. 1 shows the bone micro-architecture for healthy bone and osteoporotic bone.

Healthy Bone

Osteoporotic Bone

Fig. 1: Bone micro-architecture

AC C

EP

Osteoporosis is more common in women than in men, because generally women have smaller and thinner bones than men and also due to the hormonal changes that occur during menopause. The conventional technique used to measure bone mineral density (BMD) is Dual-Energy X-ray Absorptiometry (DXA), which emits an extremely low dose of ionizing radiation compared to radiographs. The results of texture analysis of the bone structure on radiographs have been compared to BMD in several studies [13]. Conventional bone radiographs are considered as a cheaper alternative to DXA. Moreover, it has been shown that texture analysis of X-ray radiographs is well suited to assess changes in trabecular bone architecture [41, 42].

The aim of the approach proposed here is to complete the diagnosis achieved in clinical routine based on the BMD measured by DXA, by characterizing and quantifying the quality of the bone micro-architecture. The main issue is how to improve the early diagnosis of osteoporosis and assess bone quality parameters on Xray radiographs. The calcaneus was used because of the limited amount of soft tissues surrounding this bone. Soft tissues could increase the variability of the method. Moreover, the calcaneus consists of 90% trabecular bone and is a good predictive site

2

ACCEPTED MANUSCRIPT

AC C

EP

TE D

M AN U

SC

RI PT

of fracture in terms of bone mineral density [14]. Currently, X-ray is a widely used modality in the hospital environment. Quantitative Ultrasound (QUS) are also used to predict fragility fractures in postmenopausal women. Although conventional QUS approaches confer numerous advantages such as absence of ionizing radiation, portability, lower costs, availability in primary care, they do not provides added value compared to DXA for osteoporosis diagnosis and cannot be used to monitor the skeletal effects of treatments for osteoporosis [40]. Updated official positions of the International Society for Clinical Densitometry [43] validated only the heel as a skeletal site for the clinical use of QUS in osteoporosis management. However recent advanced quantitative echo sound methodology at the proximal femur might be promising [44]. In this section, relevant work on the image processing based diagnosis of osteoporosis is summarized. Geraets et al. [30] developed a method for the automated recognition of the radiographic trabecular pattern for osteoporosis diagnosis. Genant et al. [5] proposed an imaging assessment of bone quality in osteoporosis by computed tomography (CT) images and micro-computed tomography (µ-CT) images. E. Legrand et al. [6] showed the analysis of trabecular bone micro-architecture for measurement of bone mineral density and vertebral fractures for diagnosis of osteoporosis. This study was done for diagnosis of osteoporosis in male subjects and reported results were interesting and motivating. J. S. Gregory et al. [7] analyzed the trabecular bone structure using Fourier transforms and artificial neural networks for diagnosis of osteoporosis. M. G. Roberts et al. studied the image texture in dental panoramic radiographs for diagnosis of osteoporosis [8]. M.S. Kavitha et al. [9] developed computer aided system for diagnosis of osteoporosis from dental panoramic radiograph images. Support Vector Machine (SVM) classifier was used to separate osteoporotic patients from healthy subjects. M. Ascenzi et al. [10] studied the variation of trabecular architecture in proximal femur of postmenopausal women for diagnosis of osteoporosis. B. R. Gomberg et al. [11] explained topological analysis of trabecular bone MR images for diagnosis osteoporosis. V. Sapthagirivasan et al. [12] presented a computer vision based diagnosis system for osteoporotic risk detection using an SVM classifier for X-ray digital images. Haidekker et al. [31] proposed an image processing based method for the diagnosis of osteoporosis and estimation of the individual bone fracture risk. R. Dendere et al. [32] developed a model to measure phalangeal bone mineral mass on a slot-scanning digital radiography system. Bayarriet et al. [33] presented a mechanical characterization of trabecular bone applied to MRI examinations for osteoporosis diagnosis. Cheng et al. [34] proposed learning based landmark detection for osteoporosis analysis. E. Hassouni et al. classified trabecular bone X-ray images using fractional Brownian motion and the Probability Density Function. T. Whitmarsh et al. [35] developed a method for reconstruction of 3 dimensional shape and distribution of bone mineral density of proximal Femur from Dual-Energy X-Ray absorptiometry. V. Sapthagirivasan et al. [36] analyzed texture pattern of femur X-ray images for diagnosis of osteoporosis. Texture analysis was also done by Gaidel et al. [37] for automated diagnosis of osteoporosis by plain hip radiography. J. H. Potgieter et al. [38] developed a method to measure bone mineral density on a slot-scanning digital radiography system using dual energy X-ray absorptiometry. H. H. Lin et al. [39] developed a novel two

3

ACCEPTED MANUSCRIPT

component model to calculate bone mineral density and volume fractions from computed tomography images.

RI PT

Although all these studies are interesting and the experimental results reported are encouraging, there is still a need to design accurate, computer vision based systems for the diagnosis of osteoporosis from bone images.

M AN U

SC

The main contribution of this paper is a novel and efficient imaging method to improve the diagnosis of osteoporosis based on strategically extracted discriminatory features from bone X-ray images. Using these discriminatory features, supervised classification techniques are used to distinguish osteoporotic patients from healthy subjects automatically. The only manual intervention required is the specification of a measurement region within the image. The intensity and contrast of the bone X-ray images are adjusted prior to feature extraction, which ensures that the extracted features efficiently discriminate between the two populations. These features are subjected to different classifiers, resulting in an accurate classification in comparison to features extracted without pre-processing. Another highlight of the proposed work is the low time complexity required for the proposed method. Discriminatory features extracted from bone X-ray images are standardized using the z-score normalization technique and compressed using principal component analysis which ensures low time complexity.

2. Materials

TE D

The remaining part of the work is organized as follows. Section 2 presents the materials used in this work. Section 3 describes the proposed image processing based method including pre-processing, feature analysis and classification. Section 4 reports and discusses the experimental results obtained by the proposed algorithm. Section 5 concludes the paper.

AC C

EP

This section describes the protocol for selecting the clinical population and the radiographic process for data acquisition. 2.1 Populations This is a retrospective multi-centric study that was conducted at three French research hospitals (Amiens, Orleans and Cochin in Paris). Patients were recruited from a program originated by the hospital of Orleans between Nov. 2004 and Feb. 2006. The study involved 174 women aged 40 to 92 years attending the bone densitometry unit and hospitalized in the rheumatology, orthopedic, and geriatric units, and comprised 87 controls (CC) and 87 patients with osteoporotic fractures (OP). Among the OP subjects, there were 21 patients with hip fracture (HF), 22 with vertebral fracture (VF), 23 with wrist fracture (WF) and 21 with other fractures (OF). All OP patients (aged 71.34 ± 10.55 SD) and control subjects (aged 68.93 ± 9.78 SD) filled out an osteoporosis risk questionnaire (age, family history of fracture, menopausal status, use of tobacco or alcohol, use of hormonal replacement therapy (HRT), treatment by oral corticosteroids, rheumatoid arthritis). Patients treated with corticosteroids, fluoride, bisphosphonates, HRT, tibolone, calcitonin, SERM, and PTH for more than 6 months in the previous year were excluded from the study.

4

ACCEPTED MANUSCRIPT

AC C

EP

TE D

M AN U

SC

RI PT

Patients who were immobilized for more than 2 months and with known diseases that could interfere with bone metabolism were also excluded from the study. All fracture cases were reviewed by experienced investigators who considered the diagnosis of fragility fracture if it occurred after the age of 40 years. The cases were described as either spontaneous fractures, fractures resulting from strenuous activity, fractures after falls from standing height or less (low trauma energy) and following radiologic data. The presence of vertebral fractures was diagnosed using lateral spine radiographs according to Genant’s semi-quantitative classification with grade II or more. BMD was performed by DXA on all patients and controls. Grade I vertebral fractures were not included in this study since in pivotal OP treatment trials only grades II or III were considered. However, OP characterized by wrist fractures (23 cases in the present study) was included, as they can be considered to represent an early stage of osteoporosis. Since these fractures occur in patients younger than those with hip or vertebral fractures. 2.2 Acquisition Images were obtained on the calcaneus (bone of the heel) with a direct digital Xray device (BMA™, D3A Medical Systems) [13].The calcaneus is a bone surrounded by limited soft tissues and it was chosen to limit the effect of soft tissues that could increase the variability of the method. Radiographic images were captured following a highly standardized protocol. The same parameters were used for all patients: focal distance (1.15 m), X-ray parameters (55 kV and 20 mAs). Using two anatomical landmarks as previously described in [14], a similar Region of Interest (ROI) was defined for each subject. An operator localized these land marks on each image (Fig. 2(a)) enabling the software device to position the ROI (1.6×1.6 cm2) (Fig. 2(b)). Images were coded into 16 bits (400×400 pixels size). Fig. 2(b) and Fig. 2(c) show respectively a representative image of a ROI of a control subject and of an osteoporotic patient.

(a)

(b)

(c)

Fig.2: X-ray of a calcaneus bone with the ROI (a), segmented ROI of a control subject (b), extracted ROI of an osteoporotic patient (c).

5

ACCEPTED MANUSCRIPT

3. The proposed Imaging Method

SC

RI PT

The main objective of this study was to design and develop an image processing based method to distinguish osteoporotic patients from healthy subjects. The proposed method is based on discriminatory features extracted from pre-processed bone X-ray images that are subjected to artificially intelligent classifiers. To extract efficient and discriminatory features, images were pre-processed to improve their quality before feature extraction. The extracted features were normalized and compressed to reduce the time complexity and to improve classifier performance. The proposed method is broadly divided into four steps: pre-processing, feature extraction and analysis, prominent feature selection using supervised classification. Fig. 3 explains the flowchart of the proposed method. Bone X-ray Image Acquisition

Machine learning/Computer Vision

X-ray Machine Set up

M AN U

Quality Enhancement/ Preprocessing of X-ray Image

X-ray of a calcaneus bone with the ROI

Testing Samples

Supervised Classification

Database

Training Data

EP

Healthy & Osteoporotic patient X- ray Image

AC C

p- value Calculation to select most Discriminatory Features

Most Discriminatory Feature Selection (p-value ≤ 10-6 )

TE D

Identification/Grading of Healthy & Osteoporotic patient X- ray Image

Statistical/Texture Feature Extraction and Analysis

Validation of Results with Doctor’s grading

Testing Samples

Classification Healthy/ Osteoporotic patient X- ray Image

Performance Measure

Accuracy

Time Complexity

Classification of two populations Healthy/osteoporotic

Fig. 3: The flowchart of the method using X-ray images of calcaneus bone.

6

ACCEPTED MANUSCRIPT

Step 1: Pre-processing

TE D

M AN U

SC

RI PT

X-ray images of the bone micro architecture from osteoporotic patients and healthy subjects show high similarity, making the diagnosis of osteoporosis from X-ray images a challenging task. To address this issue, each bone X-ray image was preprocessed to enhance its quality before feature extraction by adjusting the intensity and contrast of the input image. Intensity adjustment was done by transforming the intensity values to new values in such a way that 1% of data was saturated at lower and higher intensity values in an input image [15]. Contrast enhancement was accomplished using adaptive histogram equalization [15]. Fig. 4 (a) shows two samples of bone X-ray images from healthy subjects. Fig. 4(b) and 4(c) present the images after intensity adjustment and contrast enhancement, respectively. Similarly, Fig. 5(a) shows two samples of X-ray images from osteoporotic patients. Fig. 5(b) and Fig. 5(c) present the images after intensity and contrast adjustment, respectively. It can be seen that the quality of the images has been changed, and that trabeculae are more visible. The pre-processed images were further used for feature extraction.

AC C

EP

(a) (b) (c) Fig. 4: Two representative samples of bone X-ray images of healthy subjects. original input images (a), images after intensity adjustment (b), images after contrast adjustment(c).

(a) (b) (c) Fig. 5: Two representative samples of bone X-ray images of osteoporotic patients (with fracture). original input images (a), images after intensity adjustment (b), images after contrast adjustment (c).

7

ACCEPTED MANUSCRIPT

Step 2: Feature Extraction and Analysis

RI PT

The density of bone and its micro-architecture depend on the composition of calcium and collagen. In osteoporotic patients, the steady loss of collagen and calcium leads to a decrease in bone density and deterioration of bone micro-architecture which may disturb the pattern of pixel intensities in bone X-ray images. First-order Statistical Features: First-order statistical features characterize the distribution of the image intensities across an entire image. The mean and standard deviation of all pixel intensities within the ROI in each image were used.

M AN U

SC

Second-order Statistical Features: In osteoporotic patients, insufficient bone formation may cause changes in the texture of bone X-ray images. These changes will be reflected in the relationships between pixel intensities and their distribution over the image. In contrast to first‐order statistical features, second‐order features of an input image in the spatial domain quantify relationships between the intensities of neighboring pixels. Second-order statistical features can be calculated using a gray-level co-occurrence matrix (GLCM). A GLCM is a second-order statistical textural analysis, which shows the relationship between intensity values of neighborhood pixels in a gray image [16]. For an image, I of size R×C, the GLCM matrix, G, can be calculated [16] using the following equation: Gd (i,j) = |{(, ), ( + ,  + ) ∶ (, ) = , ( + ,  + ) = }| (1)

TE D

where (r,c), (r+ ,  + ) ∈R×  i = (, )and j = ( + ,  + ) d = ( , ) and |·| represents the cardinality of a set.

For a pixel intensity value i in the image, the probability of a pixel intensity value j at the ( , ) distance is:  (,)

Pd (i ,j) = ∑ 

 (,)

(2)

EP

Using equations (1) and (2), four texture features: energy, contrast, homogeneity and correlation can be calculated using the mathematical formulas given below: Energy = ∑ ∑[ (, )]!

(3)

AC C

Contrast= ∑ ∑ [( − )! { (, )}] (4)

Homogeneity = ∑ ∑

Correlation =∑ ∑

# (,)

(5)

$%(&)'

(&µ ) *+&µ , -# (,)

(6)

.) .,

A total of six statistical features were extracted and studied from the bone X-ray images before and after pre-processing to discriminate between healthy and osteoporotic patients. To check the discriminating behavior of the extracted features, a p-value test was calculated using the rank sum Wilcoxon test, also called the MannWhitney U test [18, 19].

8

ACCEPTED MANUSCRIPT

RI PT

The rank sum Wilcoxon test is a non-parametric statistical hypothesis test which uses a z‐statistic to calculate the p-value. This calculated p-value is used to determine if two sets of data are significantly different from each other. A lower p-value indicates higher discrimination and a larger p-value shows that there is no discrimination between two populations for that feature. On the basis of the calculated p-value, only the most prominent features were selected to distinguish osteoporotic patients from healthy subjects.

Step 3: Prominent Feature Selection

M AN U

SC

Intelligent classifiers based on machine learning were used to separate osteoporotic patients from healthy subjects. As the performance of the classifiers improves if the range of features subjected to the classifier is standardized, z-score normalisation [20] was applied to each feature. It is well known that the accuracy of the classifiers depends on the discrimination of features used for the classification. If the features subjected to the classifier are more prominent, then the classification accuracy is better. On the other hand, the complexity of the classifier will be reduced if fewer features are subjected to the classifier. In this study, a dimension reduction technique was used to reduce the feature vector by considering only discriminatory features. Dimension reduction techniques compress the high dimensional feature vector by assigning higher weight to discriminatory features. On the basis of experiments, it was found that the dimension reduction algorithm Principal Component Analysis (PCA) [21, 22] provided the most prominent features for the set of features used here.

EP

TE D

PCA transforms the observation data of possibly correlated variables into a set of values of linearly uncorrelated variables on the basis of orthogonal transformation. These transformed linearly uncorrelated variables are known as principal components (PCs). PCA maps the observation data in such a way that the first PC has a higher discrimination, the second has a lower discrimination than the first PC, and so on. Here, only discriminatory features were subjected to the classifiers. The p-value was calculated for all principal components after PCA reduction to select prominent features for classification as shown in the results below. TABLE 1 Statistical features and corresponding p– values (before pre-processing)

AC C

Features

Healthy subjects (mean ± std. deviation)

Osteoporotic patients (mean ± std. deviation)

p- values

(1.94E+003) ± 66.22 (272.73) ± 20.49

0.0626 0.1449

(2.70E+004) ± 256.5899 -0.02 ± 0.01 (6.2936e-006) ± (2.16E-008) 0.03 ± (1.72E-004)

0.5160 0.2309 0.4675 0.4563

First -order statistical feature: Mean Standard Deviation

(1.99E+003) ± 104.69 (289.39) ± 39.42

Second-order statistical features: Contrast Correlation Energy Homogeneity

(2.69E+004) ± 249.9562 -0.01 ± 0.01 (6.29E-006) ± (1.62E-008) 0.02 ± (1.81E-004)

9

ACCEPTED MANUSCRIPT

First -order statistical feature: Mean Standard Deviation

(3.27E+004) ± 50.75 (1.69E+004) ± 122.81

(3.27E+004) ± 90.22 (1.67E+004) ± 94.18

Second-order statistical features: (2.73E+004) ± 796.11 -0.05 ± 0.03 (7.92E-006) ± (1.39E-007) 0.02 ± (6.13E-004)

(2.76E+004) ± 689.29 -0.06 ± 0.03 (7.8953e-006) ± (1.63E-007) 0.02 ± (5.68E-004)

0.0020 1.21E-011

0.4735 0.2748 0.6000 0.7839

SC

Contrast Correlation Energy Homogeneity

p- values

RI PT

TABLE 2 Statistical features and corresponding p – values (after pre-processing) Healthy subjects Osteoporotic patients Features (mean ± std. deviation) (mean ± std. deviation)

TABLE 3 Features (Principal components) and their corresponding p-values (after reduction using PCA)

PC 1 PC 2 PC 3 PC 4 PC 5 PC 6

Healthy subjects (mean ± std. deviation) (-4.02E-006) ± (6.47E-006) (-1.80E-006) ± (4.06E-006) (-2.62E-008) ± (1.10E-007) (-1.63E-012) ± (8.32E-011) (-4.79E-012) ± (1.64E-011) (7.84E-014) ± (2.78E-012)

Osteoporotic patients (mean ± std. deviation) (4.01E-006) ± (6.85E-006) (1.80E-006) ± (2.83E-006) (2.61E-008) ± (1.96E-007) (1.63E-012) ± (9.52E-011) (4.79E-012) ± (2.40E-011) (-7.82E-014) ± (2.63E-012)

M AN U

Features

p- values

1.69E-006 2.24E-005 0.0013 0.8663 0.6834 0.8436

AC C

EP

TE D

Table 1 presents the results of feature analysis before pre-processing of the bone Xray images, showing the mean ± standard deviation and p-value for healthy subjects and osteoporotic patients. It can be seen that the p-value obtained is less significant for classification compared to the p-value obtained after pre-processing. Table 2 presents the results expressed as mean ± standard deviation and p-value for feature extraction after preprocessing. The lower p-value shows that the features extracted after pre-processing are more discriminative in comparison to those extracted before pre-processing. Table 3 presents the results of mean ± standard deviation and p-values for all six principal components after PCA reduction. It can be clearly seen in Table 3 that the first two principal components have much lower p-values, indicating a higher discrimination between osteoporotic patients and healthy subjects. Tables 1, 2 and 3 show the information content of the various features considered for classification. These data indicate that limited information is present in the secondorder features, and most of the information for classification is present in the mean and standard deviation of the intensities. These results also demonstrate that the amount of information present significantly increases by the pre-processing, and this has by far the greatest effect on the standard deviation, making it the most discriminatory feature for classification. Table 4 presents the linear combinations of the original variables that generate the principal components PC1 and PC2 which were used for classification in the proposed work. It can be clearly seen in Table 4 that texture feature: Contrast dominates first principal component PC1 and statistical feature: Std. deviation dominates second principal component PC2.

10

ACCEPTED MANUSCRIPT

RI PT

TABLE 4 Linear combinations of the original features that generate the principal components PC1 and PC2 Principal Components (PCs) PC1 PC2 -0.5160 -0.4609 -0.3014 0.8533 0.0268 0.8017 0.0052 -0.1397 0.0052 -0.1397 0.0052 -0.1397

Features Statistical Features

SC

Texture Features

Mean Std. deviation Contrast Correlation Energy Homogeneity

M AN U

Fig. 6 shows the box-and-whisker plots (median and the 25% and 75% quartiles) of the mean, standard deviation and GLCM features after pre-processing from healthy subjects and osteoporotic patients. As can be seen, these features have a limited capability to discriminate between healthy and osteoporotic subjects. Fig. 7 shows the box plots (mean ± standard deviation) of the first two principal components after PCA reduction. It is clearly visible in Figure 7 these two principal components can be subjected to classifiers. 4

4

x 10

x 10

1.71

3.32

Standard Deviation

1.7

3.3 3.29

3.27 3.26

0.026

0.025 0.0245

AC C

Homogeneity

0.0255

1.68 1.67 1.66 1.65 1.64

Osteoporotic Patients

EP

Healthy Subjects

0.024

Osteoporotic Patients

Healthy Subjects

Osteoporotic Patients

0 -0.02 -0.04 -0.06 -0.08

0.0235

-0.1

0.023

-0.12

Healthy Subjects

Healthy Subjects

0.02

Correlation

3.28

1.69

TE D

Mean

3.31

Osteoporotic Patients

Fig. 6: Box plots of the extracted features from healthy subjects and osteoporotic patients. (before applying PCA)

11

ACCEPTED MANUSCRIPT

-5

-6

x 10

x 10 10

1 0 -1

0

-5

Osteoporotic Patients

Healthy Subjects

Osteoporotic Patients

SC

Healthy Subjects

5

RI PT

2

Principal Com ponent 2

P rinc ipal Com ponent 1

3

Fig.7: Box plot of the two PCs subjected to the classifiers (after reduction using PCA).

M AN U

Step -4: Supervised classification of the two populations

The discriminatory feature vectors (first and second PCs) were used to separate healthy X-ray images from osteoporotic ones using different classifiers. The four most popular classification algorithms, namely Support Vector Machine (SVM) [23, 24], Naive Bayes classifier [25], Artificial Neural Network (ANN) [26] and k-Nearest Neighbors (k-NN) classifier [27], were applied to the first two principal components and the classifier that returned the best performance was finally selected for classification. Table 5 shows the best configuration settings used for the classification in terms of the results obtained.

TE D

TABLE 5 Design parameters used for classification.

Classifiers

Design parameters used for classification

Support Vector Machine (SVM) Naïve Bayes

RBF Kernel with Sigma =3 Laplace Correction – true Hidden sizes= 4 Training function: trainlm Number of nearest neighbors = 5 Distance = Cosine, Rule = Nearest

Artificial neural network (ANN)

EP

k- Nearest Neighbors (k-NN)

AC C

The SVM classifier [23, 24] was trained using a training set that included bone X-ray images from healthy subjects and osteoporotic patients. The performance of the classifier was checked for different kernels and the best parameter was finally selected to classify the test samples of bone X-ray images. Fig. 8 shows the classification process using the SVM.

12

ACCEPTED MANUSCRIPT

Test Sample Bone X-ray Images Kernels K (S1,St)

Machine Learning

S2

K (S2,St)

+

Sn

K (Sn,St)

(Healthy/Osteoporotic)

M AN U

Support Vectors from Training database of bone X-ray images

Test Sample Bone X-ray image is classified as Healthy/Osteoporotic

SC

S1

RI PT

St

Fig.8: Support Vector Machine (SVM) classification.

Training Set Bone X-ray images (Healthy/ Osteoporotic Patients)

Bayes Rule (Normal Distribution)

Find the probability of test sample to belong in possible classes (Healthy/Osteoporo tic Patients)

Best Prediction for test Sample (Healthy/Osteoporotic Patients)

EP

Test samples of X-ray images

TE D

The Naïve Bayes classifier [25] works on the Bayes rule and probability measurement. The classifier was trained using the training dataset of bone X-ray images from healthy subjects/osteoporotic patients and the best parameter setting was selected. Fig. 9 presents the flow diagram of the Naïve Bayes classification algorithm used for machine learning.

AC C

Fig.9: Bayes rule based classification.

The main parameter settings for artificial neural network classification [26] are hidden sizes and training function. The performance of the artificial neural network was checked for different parameters on the basis of the training database and the best parameter was finally selected for classification of the test images. Fig. 10 shows the process of the artificial neural network used for classification of healthy and osteoporotic patients.

13

ACCEPTED MANUSCRIPT

Hidden Layer

RI PT

Hidden Layer X1 X2

Xn Input Samples Healthy/Osteoporotic Patients from Bone X-ray Image

SC

Output Classification of Healthy/Osteoporotic Patients from Bone X-ray Image

M AN U

Neurons

Fig.10: Artificial Neural Network (ANN) classification.

In the nearest neighbors based classification algorithm [27], the number of nearest neighbors and distance metrics were selected according to the training dataset. The performance of the k-NN classification algorithm is highly dependent upon the number of nearest neighbors used so the best parameter setting was selected for classification of the test samples. Fig. 11 presents the flow of the k-NN classification algorithm used to separate healthy and osteoporotic patients.

TE D

Parameter Selection

Find the k nearest neighbors to the Test Sample of X-ray image

No. of Nearest Neighbors, Type of Distance

EP

Training Data Base Bone X-ray Image (Healthy/Osteoporotic)

Set Maximum Label Class of K to Test Sample

Classified Test Sample

AC C

Test Sample Bone X-ray Image

Fig. 11: k‐Nearest Neighbor (k-NN) classification.

4. Experimental Results and Discussion In this study, a database of 174 bone X-ray images was used for experiments. This section reports and discusses the results. 80 images out of 174 were used to train the classifiers and the remaining 94 images were used to check the performance of each classification method. There was no overlap between the images used for training and testing. Table 6 shows the distribution of healthy and osteoporotic images used for training and testing. Experiments were performed with four different classifiers to

14

ACCEPTED MANUSCRIPT

SC

RI PT

classify osteoporotic patients and healthy subjects and the classifier performances were evaluated using sensitivity, specificity and accuracy [29].The performance of all four classifiers was measured on the original features and results are reported in Table 7. It can be clearly seen in Table 7 that the original features were not significant enough for the efficient classification of the two populations. Table 8 shows the performance of the different classifiers on the discriminatory features selected after PCA. Table 8 compares the performance of the classifiers with/without pre-processing before feature extraction. As can be seen in Table 6 and Table 8, the performances of all four classifiers on the original features are not significant whereas their performance with the use of PCA is much better. All the classifiers provide an accuracy of more than 95%, which can be considered as encouraging results.

Bone X-ray image

M AN U

TABLE 6 Bone X-ray image database considered in this study. Total No. of samples used to train the classifiers 40 40 80

Healthy Subjects Osteoporotic Patients Total

Total No. of samples used to test the performance of classifiers 47 47 94

Total samples 87 87 174

TABLE 7 Performance of the different classifiers using the original features.

SVM Naive Bayes k-NN ANN

True Positive 45 45 42 40

False Positive 24 44 16 26

True Negative 23 3 31 21

TE D

Classifiers

False Negative 02 02 05 07

Sensitivity (%) 95% 95% 89% 85%

Specificity (%) 49% 7% 66% 44%

Accuracy (%) 72% 51% 77% 64%

EP

TABLE 8 Performance of the different classification methods and comparative performance with/without preprocessing before feature extraction True Positive

AC C

Classifiers

SVM Naive Bayes k-NN ANN

SVM Naive Bayes k-NN ANN

29 25 27 33

False Positive

True Negative

False Negative

Sensitivity (%)

Specificity (%)

Accuracy (%)

Feature extraction without pre-processing 23 23 19 17

24 24 28 30

18 22 20 14

61.70% 53.19 % 57.44 % 70.02 %

51.06 % 51.06 % 59.57 % 63.82 %

56.38 % 52.12 % 58.51% 67.02 %

95.74% 93.61 % 97.87 % 95.74 %

97.87% 95.74% 96.80 % 96.80 %

Feature extraction after pre-processing 47 46 45 46

02 03 01 02

45 44 46 45

15

00 01 02 01

100% 97.87 % 95.74 % 97.87 %

ACCEPTED MANUSCRIPT

x 10

RI PT

Healthy (training) Healthy (classified) Osteoporotic (training) Osteoporotic (classified) Support Vectors

1

0.5

SC

0

-0.5

-1 -3

-2

M AN U

Second Principal Component after PCA Reduction

-5

1.5

-1

0

1

2

First Principal Component after PCA Reduction

3 -5

x 10

Fig. 12: Classification results using SVM based classification

TE D

The SVM classifier shows the best performances, with 97.87% accuracy, 100% sensitivity and 95.74% specificity. Fig. 12 presents the plot of the two main discriminatory PCs for the training and testing set of SVM based classification. It is clearly visible in Fig. 12 that the two main PCs which were selected for classification are sufficiently discriminatory for accurate classification.

EP

Fig. 13 presents the ROC curves for the four classifiers (SVM, k-NN, Naïve Bayes and ANN). The area under the ROC curve (AUC) and Youden’s index were also estimated from the ROC. Values are reported in Table 9. The highest AUC value was 0.9954 for the k-NN classifier and the highest Youden’s index was 0.9574 for the SVM classifier

AC C

Classifiers SVM Naive Bayes k-NN ANN

TABLE 9 ROC curve analysis Area under curve (AUC) 0.9824 0.9791 0.9954 0.9681

Youden’s index 0.9574 0.9361 0.9149 0.9361

The algorithms were implemented in MATLAB R2011b (Math Works) software using CPU@ 2.3 GHz, 4GB RAM, 32-bit operating system. The computational time needed for feature extraction was approximately 9 seconds per image and for feature normalization, reduction and classification approximately11 sec per image, which can be considered as very low. The low time complexity of the proposed method makes it suitable for classification of the trabecular bone structure of osteoporotic Patients in real time

16

ACCEPTED MANUSCRIPT

ROC

ANN Classifier

0.95

SVM Classifier k-nn Classifier

RI PT

1

0.9

0.85

SC

True Positive Rate

Naive Bayes Classifier

0.75

0

0.05

M AN U

0.8

0.1

0.15

0.2

0.25

False Positive Rate

Fig. 13: ROC curve for SVM, k-NN, Naïve Bayes and ANN classifier

Leave One Out Cross Validation (LOOCV)

TE D

The leave one out cross validation (LOOCV) technique [28] was used to distinguish osteoporotic patients from control subjects. In this method, only one sample was used for testing purposes and the remaining samples were used to train the classifier. This procedure was repeated for the complete dataset. The LOOCV technique was used for SVM as it provided the best classification performance. Table 10 presents the crossvalidation results. The LOOCV gives 96.55% accuracy with 95.40% specificity and 97.70% sensitivity with the SVM classifier, indicating that the proposed approach is efficient and may be helpful for the diagnosis of osteoporosis in real time.

EP

TABLE 10 Results obtained using the SVM classifier with the leave one out cross validation. Support Vector Machine Kernel: RBF with σ = 3 174 173 1 87 87 85 02 83 04 97.70 % 95.40 % 96.55 %

AC C

Classifier used Classifier setting parameters Total number of samples used Samples used for training Samples used for testing Total number of osteoporotic subjects Total number of control Subjects True positive (TP) False negative (FN) True negative (TN) False positive (FP) Sensitivity Specificity Accuracy

17

ACCEPTED MANUSCRIPT

TABLE 11: Comparison between published studies and the proposed approach (data taken from the literature)

M.S. Kavit ha et al. [9]

V. Sapthagiriv asan et al. [12]

A. Gaidel et al. [37]

Imaging method for the diagnosis of osteoporosis

Imaging analysis for osteoporosis diagnosis

Dental panoramic radiographs (DPRs) from 663 female patients

Extraction of texture feature from DPR images and classification

Dental panoramic radiographs from 100 female subjects

SVM classifier for continuous measurement of the cortical width of the mandible on dental panoramic radiographs.

Performance measurements

Time complexity

RI PT

CT images, QCT images µ-CT image

Not reported

Area under curve: 0.872

Sensitivity: 90.9% specificity: 83.8%

Not reported

Not reported

Average classification time: 9 sec per image

hip radiographs images of 50 subjects

trabecular features extraction and SVM classification

Mean accuracy: 90%

50 sec for feature extraction and 8 sec per image for classification

Plain hip radiograms of 42 patients

Texture analysis of plain hip radiograms

Diagnostic error probability 0.2

Not reported

Preprocessing, statistical feature Analysis, feature normalization & reduction, discriminatory feature Selection, classifiers for osteoporosis diagnosis

Accuracy: 97% AUC (area under curve) = 98.24%

20 Sec

Bone X-ray image 87 healthy subjects and 87 patients with osteoporotic fractures

EP

Proposed study

Imaging assessment of bone quality in osteoporosis Diagnosis of osteoporosis using image texture in Dental Panoramic Radiographs Diagnosis of osteoporosis from dental panoramic radiographs Diagnosis of osteoporosis by extraction of trabecular features from hip radiographs Texture analysis for diagnosis of osteoporosis by plain hip radiography

Methodology used

SC

M. G. Roberts et al. [8]

Database

M AN U

Genant & Jiang [5]

Objective

TE D

Reference

AC C

Table 11 compares different existing studies for the diagnosis of osteoporosis. We are aware that the datasets used in these various studies are not the same and that confrontation of the different methods should be done on the same database. Since we have no access to these data, this task remains difficult. The proposed method achieved 97% accuracy, which is encouraging. However, this study has some limitations. The proposed method has only been validated on a small database. The bone of the calcaneus that was used is not commonly used in the clinical routine. It is necessary to be aware that to complete the DXA diagnosis employed in the clinical routine, our proposed method will increase the cost of the diagnosis and augment the radiation dose as well as the duration of the whole examination.

18

ACCEPTED MANUSCRIPT

5. Conclusion

M AN U

SC

RI PT

This work has proposed a novel and efficient imaging method for osteoporosis diagnosis using bone X-ray images. Statistical and texture features strategically extracted from pre-processed X-ray images were used to discriminate between healthy subject and osteoporotic patient images. Extracted features normalized using the z-score method and reduced using PCA were subjected to machine learning and classification. Four different classifiers (SVM, Naïve Bayes, k-NN and ANN) were used to discriminate between the two populations. Results demonstrate that all the classifiers achieved accuracy higher than 95%. The SVM classifier provided the best performances with an accuracy of 97.87%, sensitivity of 100% and specificity of 95.74%, which is encouraging. The proposed diagnostic system shows a higher accuracy with lower time complexity in comparison with existing methods for the bone tissue characterization. Future research work in this area might be to explore other discriminatory features like orientation-dependent features and correlation of first-order features with BMD and to use different classifiers to obtain even greater accuracy with less time complexity.

References:

D. Marshall, O. Johnell, H. Wedel, Meta-analysis of how well measures of bone mineral density predict occurrence of osteoporotic fractures, BMJ. 312 (1996) 1254-1259. [2] J. A Kanis, P. Delmas, P. Burckhardt, C. Cooper, D. Torgerson, Guidelines for diagnosis and management of osteoporosis: The European Foundation for Osteoporosis and Bone Disease, Osteoporosis International. 7 (1997) 390–406. [3] Mithal A, Kaur P. Osteoporosis in Asia: a call to action. Curr Osteoporos Rep. 2012 (4), 245– 247. [PubMed]. [4] S. A. Goldstein, The mechanical properties of trabecular bone: dependence on anatomic location and function, J. Biomech. 20 (1987) 1055-1061. [5] H. K. Genant, J. Y. Jiang, Imaging assessment of bone Quality in osteoporosis, Clinical Reviews in Bone and Mineral Metabolism, Springer. 4(2006) 213–224. [6]E. Legrand, D. Chappard, C. Pascaretti, M. Duquenne, S. Krebs, V. Rohmer, M. F. Basle, M. Audran, Trabecular bone micro-architecture bone mineral density and vertebral fractures in male osteoporosis, J. Bone Mineral Res., 15 (2000), 13-19. [7] J. S. Gregory, R. M. Junold, P. E. Undrill, R. M. Aspen, Analysis of trabecular bone structure using Fourier transforms and neural networks, IEEE Transactions on Information Technology in Biomedicine, 3 (1999), 289 – 294. [8] M. G. Roberts, J. Graham, H. Devlin, Image Texture in Dental Panoramic Radiographs as a Potential Biomarker of Osteoporosis, IEEE Transactions on Biomedical Engineering, 60 (2013), 2384 – 2392. [9] M.S. Kavitha, A. Asano, A. Taguchi, Diagnosis of osteoporosis from dental panoramic radiographs using the support vector machine method in a computer-aided system, BMC Med. Imaging, 12 (2012), 1-11. [10] M. Ascenzi, N. Hetzer, A. Lomovtsev, Variation of trabecular architecture in proximal femur of postmenopausal women, J. Biomech., 44 (2011), 248-256. [11] B. R. Gomberg, P. K. Saha, H. K. Song, S. N. Hwang, F. W. Wehrli, Topological analysis of trabecular bone MR images, IEEE Transactions on Medical Imaging, 19 (2000), 166-174. [12] V. Sapthagirivasan, M. Anburajan, Diagnosis of osteoporosis by extraction of trabecular features from hip radiographs using support vector machine: an investigation panorama with DXA, Computers in Biology and Medicine, Elsevier. 43 (2013) 1910–1919.

AC C

EP

TE D

[1]

19

ACCEPTED MANUSCRIPT

AC C

EP

TE D

M AN U

SC

RI PT

[13] E. Lespessailles, C. Gadois, G. Lemineur, J. P. Do-Huu, C. L. Benhamou, Bone texture analysis on direct digital radiographic images: precision study and relationship with bone mineral density at the osclacis. Calcified Tissue International, 80 (2007) 97–102. [14] C.L. Benhamou, E. Lespessailles, G. Jacquet, R. Harba, R. Jennane, T. Loussot, D. Tourliere, W. Ohley, Fractal organization of trabecular bone images on calcaneus radiographs, J. Bone Miner Res. 9 (1994) 1909–1918. [15] R.C. Gonzalez, R. E. Woods, Digital image processing, Prentice Hall, NJ, 2001. [16] M. M. Galloway, Texture classification using gray level run length, Comput. Graph. Image Process. 4(1975) 172–179. [17] J. H. Tan, E. Y. K. Ng, U. R. Acharya, Study of normal ocular thermogram using textural parameters, Infrared Phys. Technol, 53 (2010) 120–126. [18] J. D. Gibbons, Nonparametric, Statistical Inference. New York: Marcel Dekker, 1985. [19] M. Hollander, D. A. Wolfe, Nonparametric statistical methods. Hoboken, NJ: John Wiley & Sons, Inc. 1999. [20] M. H. Dunham, Data mining introductory and advanced topics. NJ: Prentice Hall, 2002. [21] I. T. Jolliffe, Principal component analysis series in statistics, Springer Verlag, 2002. [22] H. Abdi, L. J. Williams, Principal component analysis, Wiley Interdisciplinary Reviews: Computational Statistics, 2010. [23] C.C. Chang, C.J. Lin, LIBSVM: A library for support vector machines. 2001. http://www.csie.ntu.edu.tw/∼cjlin/libsvm. [24] S. S. Keerthi, S. K. Shevade, C. Bhattacharyya, K. R. K. Murthy, Improvements to platt’s SMO algorithm for SVM classifier design, Neural Comput.13 (2001) 637–649. [25] G. H. John, P. Langley, Estimating continuous distributions in bayesian classifiers, Proc. 11th Conf. Uncertainty Artif. Intell., San Mateo. (1995) 338–345. [26] S. Haykin, Neural networks a comprehensive foundation. 2nd edn. Pearson Education, 1999. [27] D. Coomans, D.L. Massart, Alternative k-nearest neighbour rules in supervised pattern recognition: Part 1. k-nearest neighbour classification by using alternative voting rules, 1982. Analytica ChimicaActa. 136(1982) 15–27. [28] B. Efron, R. Tibshirani, Improvements on cross-validation: The bootstrap method, Journal of the American Statistical Association, 92 (1997) 548-560. [29] E. J. Boyko, Ruling out or ruling in disease with the most sensitive or specific diagnostic test: short cut or wrong turn? medical decision making: an International Journal of the Society for Medical Decision Making. 14 (1999) 175–179. [30] W. G. M. Geraets, P. F. V. Stelt, C. J. Netelenbos, P. J. M. Elders, A new method for automatic recognition of the radiographic trabecular pattern,J. Bone Miner Res. 5 (1990) 227–233. [31] M. A. Haidekker, G. Dougherty, Medical imaging in the diagnosis of osteoporosis and estimation of the individual bone fracture risk, Biological and Medical Physics, Biomedical Engineering, Springer, 2011, 193-225. [32] R. Dendere, S. P. Whiley, T. S. Douglas, Computed digital absorptiometry for measurement of phalangeal bone mineral mass on a slot-scanning digital radiography system, Osteoporosis Int., 25 (2014), 2625-2630 [33] Bayarri, W. L. Roque, K. Arcaro, M. A. Perez, Mechanical characterization of trabecular bone applied to MRI examination: In vivo results from 100 patients, Computational vision and medical image processing, Taylor and Francis, 2014, 383-385. [34] E. Cheng, L. Zhu, J. Yang, Azhari, S. Sitam , X. Liang, V. Megalooikonomou, H. Ling, Learning based landmark detection for osteoporosis analysis, Medical Imaging, Image Processing, SPIE 9784, 2016. [35] T. Whitmarsh, L. Humbert, M. D. Craene, L. M. D. R. Barquero, A. F. Frangi, Reconstructing the 3D Shape and Bone Mineral Density Distribution of the Proximal Femur From Dual-Energy X-Ray Absorptiometry, IEEE Transactions on Medical Imaging, 30 (2011), 2101 – 2114.

20

ACCEPTED MANUSCRIPT

AC C

EP

TE D

M AN U

SC

RI PT

[36] V. Sapthagirivasan, M. Anburajan, Analysis of texture patterns in diagnosing osteoporosis using proximal femur x-ray images, Advances in Digital Image Processing and Information Technology, Springer, 205 (2011). pp. 383-392. [37] A. Gaidel, A. Khramov, Application of texture analysis for automated osteoporosis diagnostics by plain hip radiography, pattern recognition and image analysis, springer, 25 (2015), 301-305. [38] J. H. Potgieter, S. Steiner, S. P. Whiley, T. S. Douglas, R. Dendere, Dual-Energy X-Ray Absorptiometry for Measurement of Phalangeal Bone Mineral Density on a Slot-Scanning Digital Radiography System, IEEE Transactions on Biomedical Engineering, 62 (2015), 28502859. [39] H. H. Lin, S. L. Peng, J. Wu, T. Y. Shih, K. S. Chuang, C. T. Shih, A Novel Two-Compartment Model for Calculating Bone Volume Fractions and Bone Mineral Densities From Computed Tomography Images, IEEE Transactions on Medical Imaging, 36 (2017), 1094-1105. [40] E. S. Siris, Y. Chen, T. Abbott, E. Barett-Connor, P. Miller, L. Wehren, M. Berger, Bone Mineral Density Thresholds for Pharmacological Intervention to Prevent Fractures, Arch. Intern. Med., 164 (2004), 1108-12. [41] J. F. Veenland, J. L. Grashuis, H. Weinans, M. Ding, Vrooman, Suitability of texture features to assess changes in trabecular bone architecture, Pattern Recognition Letters.23 (2007), 395-403. [42] S.Majumdar, J. Lin, T. Link, J. Millard, P. Augat, X. Ouyang, D. Newitt, R. Gould, M. Kothari, H. Genant, Fractal analysis of radiographs: assessment of trabecular bone structure and prediction of elastic modulus and strength, Med Phys.26 (1999), 1333-1340. [43] B. C. Silva, S. B. Broy, S. Boutroy, J. T. Schousboe, J. A. Shepherd, W. D. Leslie, Fracture Risk Prediction by Non-BMD DXA Measures: the 2015 ISCD Official Positions Part 2: Trabecular Bone Score, Journal of Clinical Densitometry, Elsevier. 18 (2015), 309-330. [44] S. Casciaro, M. Peccaris, P. Pisani, R. Franchini, A. Greco, T. D. Marco, A. Grimaldi, L. Quarta, E. Quarta, M. Muratore, F. Conversano, An Advanced Quantitative Echosound Methodology for Femoral Neck Densitometry, Ultrasound in Medicine & Biology, Elsevier. 42 (2016), 13371356.

21