Osteoporosis assessment using Multilayer Perceptron neural networks

Osteoporosis Assessment Using Multilayer Perceptron Neural Networks Khaled Harrar 1,2,*, Latifa Hamami 1 , Sonia Akkoul 3, Eric Lespessailles 4 and Rachid Jennane 3 1 Signal

and communications laboratory, National Polytechnic School, BP 182, 16200 Algiers, Algeria [email protected]; [email protected] 2 Faculty

4 3MTO

3 PRISME

of Engineering, University M’Hamed Bougara, 35000 Boumerdes, Algeria [email protected]

institute, Hospital of Orleans, 1 rue Porte Madeleine, 45032 Orleans, France [email protected]

Laboratory, University of Orleans, 12 rue de Blois BP 6744, 45067 Orléans Cedex 2, France [email protected]; [email protected]

Abstract—The objective of this paper is to investigate the effectiveness of a Multilayer Perceptron (MLP) to discriminate subjects with and without osteoporosis using a set of five parameters characterizing the quality of the bone structure. These parameters include Age, Bone mineral content (BMC), Bone mineral density (BMD), fractal Hurst exponent (Hmean) and coocurrence texture feature (CoEn). The purpose of the study is to detect the potential usefulness of the combination of different features to increase the classification rate of 2 populations composed of osteporotic patients and control subjects. k-fold Cross Validation (CV) was used in order to assess the accuracy and reliability of the neural network validation. Compared to other methods MLP-based analysis provides an accurate and reliable platform for osteoporosis prediction. Moreover, the results show that the combination of the five features provides better performance in terms of discrimination of the subjects. Keywords—Osteoporosis, Multilayer Perceptron Networks, Cross-Validation, Texture features

Neural

I. INTRODUCTION Osteoporosis has been defined as a disease characterized by low bone mass and microarchitectural alterations of bone tissue, leading to enhanced bone fragility and consequent increase in fracture risk [1]. The most common method of assessing bone strength is to monitor loss of bone mass by Bone Mineral Densitometry (BMD) [2]. However, BMD is not the only factor involved in bone frailty and, therefore, in the individual risk of fracture [2]. Indeed, considerable overlap occurs between BMD values in patients with and without fractures [3]. Other factors that influence bone strength include the bone turnover rate, bone microarchitecture, bone mass distribution, microlesion accumulation, bone crystal quality, collagen fiber quality, the degree of mineralization, and trabecular microarchitecture [4]. A lot of work has been done to characterize the trabecular bone architecture. Jennane et al. [5] presented a series of 3D skeleton-based image processing techniques for evaluating the micro-architecture of large scale disordered porous media. The proposed hybrid skeleton method combine curve and surface thinning methods with the help of an enhanced shape classification algorithm. Results on bone demonstrate the ability of the hybrid skeleton method to provide significant topological and morphological information. Zaia et al. [6] Used Fractal lacunarity analysis to provide a model function that better represents the variation of mass density of pixels in the bone image. Clinical application suggests that the model is potentially

useful for the early diagnosis of osteoporosis. Akkus et al. [7] presented a method to determine the risk factors of osteoporosis using a multiple binary logistic regression method and to assess the risk variables for osteoporosis, in postmenopausal Turkish women. The findings of their study indicate that the use of multivariate statistical method as a multiple logistic regression in osteoporosis, which may be influenced by many variables, is better than univariate statistical evaluation. Another approach that has been used for the aim of osteoporosis assessment is Artificial Neural Network (ANN) [8]. ANNs are parallel processing structures consisting of basic processing units (neurons) which are interconnected by weighted links. ANNs have the ability of learning patterns existing in data and hencing performance classification and prediction for new data. In addition, the learning process can be supervised or unsupervised depending on whether the input data is associated with known outputs during learning or not. An ANN can have different structures based on the type of its input-output data and also its application. Among available structures; Multilayer Perceptron (MLP) has been widely used in data mining [9]. MLP is a class of feed forward neural networks which is trained in a supervised manner [9]. The aim of this study is to combine five features of bone trabecular network to discriminate 120 subjects composed of osteoporotic patients and controls. The used parameters include, Age, the bone mineral content (BMC), the bone mineral density (BMD), the hurst exponent (Hmean) and the coocurrence feature (CoEn). The BMD and BMC were measured using dual-energy X-ray absorptiometry (DXA). The Hmean which characterize the roughness of trabecular texture was estimated by the maximum likelihood estimator (MLE). Many parameters of cooccurrence exist, we selected the energy parameter (CoEn); the energy is a good predicator of the bone state. The efficiency of k-fold Cross Validation (CV) in validating the MLP outcome for osteoporosis detection is investigated. Different models of classification (SVM, Bayes Network, Multinomial logistic regression and Random Forest) are presented and compared to MLP.

II. Material and methods This section describes the protocol of the clinical study and the data acquisition. The techniques used for data mining are then presented.

A. Subjects 60 women as controls (Healthy) aged 67.9 ± 9.87 SD and 60 osteoporotic fractures cases aged 74.15 ± 10.81 SD were enrolled in the study. All the patients (Osteoporotics and controls) filled out an osteoporosis risk questionnaire that included: age, personal and familial history of fracture, tobacco (yes or no), alcohol (yes or no), menopausal status, use of hormonal replacement therapy (HRT).

B. Bone mineral measurements Bone Mineral Density (BMD) was measured using dual-energy Xray absorptiometry (DXA) (Hologic, Waltham MA, Delphi) at the hip and lumbar spine (LS) for all the subjects. The average BMD is 30.1 ± 5.19 SD for the healthy subjects and 24.49 ± 5.68 SD for the osteoporotic patients. The Bone mineral content (BMC) is derived from BMD. The average BMC of the healthy population is 0.83 ± 0.11 SD and is 0.73 ± 0.13 SD for the osteoporotic patients.

C. Image acquisition Images were obtained on calcaneus with a direct digital Xray prototype (BMA™, D3A Medical Systems, Orleans, France) [10,11]. The devices for the study were cross-calibrated. The same radiographic parameters were used for the prototypes. Focal distance was settled at 1.15 meters. X-ray parameters were 55 kV and 20 mAs for all patients. Scanning the heel permitted the selection of a similar measurement site (ROI) for each subject by using anatomical landmarks as described in [12]. These anatomical landmarks were localized by a physician, allowing a positioning of the ROI (1.6×1.6cm2) (Fig. 1). Then two texture parameters were calculated on the ROI to evaluate the bone microarchitecture quality.

A

several parallel lines in the direction under study and the obtained values were averaged to give a mean value for each direction. The Hmean value of the image was computed as the average of the measures obtained over 36 directions following 10° step of analysis [12]. The Hmean is 0.62 ± 0.02 SD for the healthy subjects and 0.60 ± 0.03 SD for the osteoporotic patients.

2) Cooccurrence parameter: A cooccurrence matrix as defined by Haralick (eq. 1) [15] is a 2D array, Ct, in which both rows and columns represent a set of possible image values. Ct(i, j) indicates how many times the gray level value, i, co-occurs following a translation vector, t = (d, ϴ), with the gray-level value, j. d represents the distance between the two pixels and ϴ, the direction. Although many parameters of cooccurrence exist, in this study, we chose to estimate the energy parameter, (eq. 1), as defined in [15]. We set d equal to 1.

CoEn   Ct (i, j ) 2

(1)

i, j

The estimation of CoEn was performed over eight directions (0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°). The mean value of the estimations was retained as the texture descriptor of each image. The CoEn is 2.81 ± 0.07 SD for the healthy subjects and 2.75 ± 0.09 SD for the osteoporotic patients. The following section describes the technique retained for the combination and the classification of the results.

E. Multilayer Perceptron Multilayer perceptron (MLP) is a feedforward neural network with one or more layers between input and output layer. Feedforward means that data flows in one direction from input to output layer (forward). This type of network is trained with the back-propagation learning algorithm [9]. MLPs are widely used for pattern classification, recognition, prediction and approximation. Multilayer Perceptron can solve problems which are not linearly separable. One input, one or more hidden and one output layer are the layers forming a MLP. Strength of each connection is expressed by a numerical value called a weight which can be updated. In general, the neuron could be modeled as a nonlinear activated function of which the total potential inputs into synaptic weights are applied [9]. A MLP with one hidden layer and its connections is illustrated in Figure 2.

B

Fig. 1. ROI for texture analysis at the calcaneus with the two anatomical landmarks A and B

D. Trabecular bone texture parameters 1) Hmean parameter: The gray levels of the ROI were fitted to the most used fractal model for texture images [13]: the fractional Brownian motion (fBm). In this case, fractal shapes are described by a single parameter, H (Hurst exponent), related to the fractal dimension, D, by H = 2 - D. The higher H is, the lower the roughness of the path is. However, due to its nonstationary nature, fBm is difficult to analyze directly. Therefore, we used the increments of fBm called fractional Gaussian noises (fGn), which are stationary processes. The maximum likelihood estimator (MLE) was applied to the fGn to estimate H [14]. The estimation of the Hmean parameter is described in detail in [12]. For each image, the H parameter was estimated on

Input Layer

Hidden Layer

Output Layer

Fig 2: The structure of a feed-forward MLP with one hidden layer An artificial neuron is the basic processing element of a neural network, which consists of a linear combiner followed by a transfer function. The neuron’s output (o) is computed by weighting the

summation of the neuron’s inputs which is then passed through a transfer function φ. This can be formulated in the Eq. 2 as:

 m  o     wi vi  bi   i 1 

(2)

where, vi is defined as the external input, m is the total number of inputs of the neuron and wi and bi are the weight and bias corresponding to the connection linking the ith input to the neuron. In this paper we have implemented the back-propagation (BP) model to classify the subjects. This algorithm is the most popular in the supervised learning architecture because of the weight error correct rules [9]. It is considered as a generalization of the delta rule for nonlinear activation functions and Multilayer networks. In a backpropagation neural network, the learning algorithm has two phases. First, a training input pattern is presented to the network input layer. The network propagates the input pattern from layer to layer until the output pattern is generated by the output layer. If this pattern is different from the desired output, an error is calculated and then propagated backward through the network from the output layer to the input layer [9].

Fig. 3. BMD, BMC and AGE for the two populations

F. Train-test and validation of the model To train and evaluate the network’s performance, we used the kfold cross validation method. In this case, the dataset is divided into k independent folds where k-1 folds are used to train the network and the remaining one is reserved for the test purpose. This procedure is then repeated until all folds are used once as a test set. The final output of the network is then computed by averaging over the obtained accuracy from each test set. We have tested different values of k. Different configurations are examined in the MLP model by changing the number of input features (AGE, BMC, BMD, Hmean, CoEn) and the number of hidden layers. The output of the model was defined as 0 for Healthy group and 1 for Osteoporotic group. The accuracy, the error and the area under curve (AUC) of the Receiver Operating Characteristic (ROC) of classification is given for each configuration. Also confusion matrices are presented for the discrimination of the subjects as follow: TN = Healthy people correctly identified FN = Osteoporotic people incorrectly identified TP = Osteoporotic people correctly identified FP = Healthy people incorrectly identified Sensitivity, Sn = TP / (TP + FN), relates to the test's ability to identify positive results. Specificity, Sp = TN / (TN + FP), relates to the ability of the test to identify negative results. Accuracy (ACC %) = (TP + TN) / (TP + FP + TN + FN), is the accuracy of discrimination of subjects.

Fig. 4. Hmean and CoEn for the two populations

III. Results The network is fed with different combination of features in each run to investigate the predictive significance of each feature. Hence, the number of input neurons is defined by the number of features and the number of hidden neurons is optimized. The network is then trained using back-propagation algorithm and validated with k-fold CV. Figure 3 illustrates a discrimination of subjects using the biometric features (BMD, BMC, and AGE). We can notice the high BMD values in the healthy subjects. The combination of the three parameters gives good results in terms of classification. As BMD and BMC are correlated, the use of one of these parameters would suffice but we preferred to present all the studied features.

Fig. 5. Accuracy for different number of neurons (K-fold = 10), using attribute selection

Table 1. Classification accuracies results Features

k-fold 10

AGE, BMC, BMD

Hmean, CoEn

All Features AS (AGE, BMC, BMD, Hmean)

ACC 86

Error 0.2

AUC 0.9

TN 52

FN 8

TP 51

20

88

0.18

0.92

52

8

30

86

0.2

0.89

51

10

77.5

0.31

0.81

20 30 10

78 75 97

0.3 0.32 0.04

20 30 10

94 93 97

20 30

96 95

FP 9

Sn 0.86

Sp 0.85

54

6

0.87

0.90

9

52

8

0.85

0.86

50

10

43

17

0.81

0.75

0.81 0.82 0.99

51 48 58

9 12 2

43 42 58

17 12 2

0.83 0.78 0.97

0.75 0.80 0.97

0.06 0.07 0.03

0.98 0.99 0.99

57 56 58

3 4 2

56 56 58

4 4 2

0.95 0.93 0.97

0.93 0.93 0.97

0.05 0.05

0.99 0.99

58 57

2 3

57 57

3 3

0.97 0.95

0.95 0.95

Table 2. A comparative of MLP to others models for k-fold = 10 and using all parameters Features MLP SVM Bayes Network Multinomial logistic regression Random Forest

ACC 97 92.5 86 96 91

Error 0.04 0.07 0.16 0.04 0.11

AUC TN FN TP FP Sn Sp 0.99 58 2 58 2 0.97 0.97 0.92 58 2 53 7 0.96 0.89 0.94 51 9 52 8 0.85 0.86 0.99 58 2 57 3 0.97 0.95 0.95 55 5 54 6 0.92 0.90 importance. Therefore, a reliable prediction system capable of the early prediction of osteoporosis is highly demanded. In the search of Figure 4 shows the result of classification of the two populations using the texture parameters (Hmean and CoEn). Low Hmean and low the best prediction models, many research studies have confirmed energy values are observed in osteoporotic patients. The outputs of ANN as a good modeling approach for osteoporosis assessment. In ACC, Sn and Sp for the designed network using different number of this regard, Lemineur et al. [8] have investigated potential role in folds for k-fold CV are illustrated in Table 1. Confusion matrices and osteoporosis of ANN. They have used 7 risk factors as inputs of ANN error classification are presented too. Different configurations are used and have obtained 78% accuracy on the test database. Gregory et al. in the network combining the different parameters. Another approach [16] presented an approach for computer analysis of trabecular bone called Attribute Selection (AS) which selects the best parameters is structure. Their techniques have used a Fourier transform to generate a investigated to improve the performance of the network. This spectral fingerprint of an image and this information passed to ANN approach evaluates the worth of a subset of attributes by considering for classification of osteoporosis and osteoarthritis. They obtained the individual predictive ability of each feature along with the degree success of classification up to 84%. of redundancy between them. Figure 5 presents the prediction accuracy obtained using 10-folds cross-validation and attribute In our study, we have used 5 factors as input for ANN and used selection approach. Two neurons in hidden Layer suffice to get the the back-propagation algorithm in the Multilayer Perceptron. We high accuracy rate (97%). trained and tested the performance network with the k-fold crossvalidation method. Considering the network accuracy, sensitivity and A comparison of the Multilayer Perceptron neural network to specificity using different k-fold CVs illustrated in Table 1, the Support Vector Machine (SVM), Bayes Network, Multinomial attribute selection approach (combining AGE, BMC, BMD, Hmean) logistic regression and Random Forest is presented in Table 2, using is preferred over the other configurations as it obtains better and more 10-folds cross-validation and combination of all parameters. As we reliable results (success reaches up to 97% for k=10). Removing CoEn have done for MLP, we have tested the performance classification of from the set of all parameters results in the high accuracy, sensitivity other models on our data using confusion matrices. As we can see the and specificity. In addition, removing AGE, BMC, BMD or Hmean best performance is up to MLP, with high accuracy classification, low from the set, results in a lower accuracy for sensitivity and specificity. error and largest AUC. The less performance is up to Bayes Network. Moreover, investigating the output results for different values of k for IV. Discussion k-fold CV shows that 10-fold CV is a better choice for network validation with the all dataset, the results are observed in table 1 where we have tested 20-fold and 30-fold CV. The AUC shows high A good deal of research conducted in the field of osteoporosis values for all configurations, which leads to think that the Multilayer detection has led to the identification of many parameters for bone Perceptron neural network constitute a reliable platform in characterization. However, besides exploring parameters, finding the classification. Moreover the combination between quantity of bone relationship between these features to those previously used along (BMD, BMC) and quality of the bone microarchitecture (Hmean) with the additional information they can provide is of great

gives the best results in terms of discrimination between subjects suffering and not from osteoporosis. Trying different number of neurons in hidden layer, we have noticed that 2 neurons are sufficient. The high accuracy is reached after a certain number of neurons in hidden layer. Increasing the number of neurons has no effect on the accuracy beyond a certain value.We have also tested the performance classification of other models: SVM, Bayes Network, Multinomial logistic regression and Random Forest. The less performance is up to Bayes Network (with accuracy = 86% and Error = 0.16). We can notice the good performance for the Multinomial logistic regression model (with accuracy = 96% and Error = 0.04). Our findings suggest the reliability of the MLP model comparing to other models.

V. Conclusion This study presents an evaluation of five features of the bone structure for the early diagnosis of osteoporosis using a MLP neural network. The main objective of the study was to investigate the neural network ability in capturing nonlinear interaction of these features for osteoporosis. We have also assessed the effectiveness of k-fold CV for MLP outcome evaluation in case of having osteoporosis dataset containing limited number of data. The results confirm the superiority of 10-fold CV especially for a limited number of data. We have reached up to 97% correct estimation. In comparison we have tested SVM, Bayes Network, Multinomial logistic regression and Random Forest methods. We only obtained 86% correct classification for Bayes Network, where the accuracy reached 96% for Multinomial logistic regression technique. Several techniques and more than 5 parameters need to be used in parallel to appreciate the pathophysiological mechanisms of osteoporotic states.

REFERENCES [1] Consensus development conference diagnosis. Prophylaxis and treatment of osteoporosis. American Journal of Medicine, 94(6): 646-550, 1993. [2] D. Marshall, O. Johnell and H. Wedel, Meta-analysis of how well measures of bone mineral density predict occurrence of osteoporotic fractures. BMJ, 312(7041): 1254–1259, 1996. [3] L. J. Melton, S. H. Kan, M. A. Frye, H. W. Wahner, W. M. O’Fallon and B. L. Riggs. Epidemiology of vertebral fractures in women. American Journal of Epidemiology, 129(5): 1000–1011, 1989. [4] R. B. Mc Creade and A. S. Goldstein. Biomechanics of fracture: is bone mineral density sufficient to assess risk?, Journal of Bone and Mineral Research, 15(2): 2305–2308, 2000. [5] R. Jennane, A. Almhdie, G. Aufort and E. Lespessailles. 3D ShapeDependent Skeletonization Method for Trabecular Bone Characterization. Medical Physics Journal, 39(1), 2012. [6] A. Zaia, R. Eleonori, P. Maponi, R. Rossi and R. Murri. MR Imaging and Osteoporosis: Fractal Lacunarity Analysis of Trabecular Bone. IEEE transactions on information technology in biomedicine, 10(3): 498 – 489, 2006. [7] Z. Akkus, H. Camdeviren, F. Celik, A. Gur and K. Nas. Determination of osteoporosis risk factors using a multiple logistic regression model in postmenopausal Turkish women. Saudi Medical Journal, 26(9): 1351-1359, 2005. [8] G. Lemineur, R. Harba, N. Kilic, O. N. Ucan, O. Osman and L. Benhamou. Efficient estimation of osteoporosis using artificial neural networks. Industrial Electronics Society, 2007. IECON 2007. 33rd Annual Conference of the IEEE. 3039 – 3044, 2007. [9] S. S. Haykin, Neural Networks and Learning Machines. 3rd Edn., Prentice Hall, New York, 2009. ISBN: 0131471392, pp: 906.

[10] E. Lespessailles, C. Gadois, I. Kousignian, J. P. Neveu , P. Fardellone, S. Kolta, C. Roux, J. P. Do-Huu and C. L. Benhamou, Clinical interest of bone texture analysis in osteoporosis: a case control multicenter study. Osteoporosis International, 19(7): 1019-1028, 2008. [11] E. Lespessailles, C. Gadois, G. Lemineur Do-Huu JP and Benhamou L. Bone texture analysis on direct digital radiographic images: precision study and relationship with Bone Mineral Density at the os calcis. Calcified Tissue International, 80(2): pp. 97–102, 2007. [12] C. L. Benhamou, E. Lespessailles, G. Jacquet, R. Harba, R. Jennane, T. Loussot, D. Tourliere and W. Ohley, Fractal organization of trabecular bone images on os calcis radiographs. Journal of Bone and Mineral Research, 9(12): 1909-1918, 1994. [13] K. Harrar and L. Hamami, The Fractal Dimension correlated to the bone mineral density”, WSEAS Transactions on signal Processing. 4(3): 110-126, 2008. [14] T. Lundahl, W. J. Ohley, S. M. Kay and R. Siffert. Fractional brownian motion: a maximum likelihood estimator and list application to image texture. IEEE, Transaction on Medical Imaging 5(3): 152-161, 1986. [15] R. Haralick, Statistical image texture analysis. In: Handbook of Pattern Recognition and Image Processing. Academic Press, San Diego, 1986, 247279 [16] J.S. Gregory, R.M. Junold, P.E. Udrill and R.M. Aspden. Analysis of trabecular bone structure using Fourier transforms and neural networks”. IEEE, Information Technology in Biomedicine, 3(4): 289-294, 1999.