INT J COMPUT COMMUN, ISSN 1841-9836 8(3):354-365, June, 2013. Breast Cancer Diagnosis based on Spiculation Feature and Neural Network Techniques V. Bălănică, I. Dumitrache, L. Preziosi Victor Bălănică, Ioan Dumitrache University Politechnica of Bucharest Romania, 060042 Bucharest, Splaiul Independentei, 313 vicord20011@gmail.com, idumitrache@ics.pub.ro Luigi Preziosi University of Politechnics Torino Italy, 10129 Torino, Corso Duca degli Abruzzi, 24 luigi.preziosi@polito.it Abstract: The degree of spiculation of the tumor edge is a particularly relevant indicator of malignancy in the analysis of breast tumoral masses. This paper intro- duces four new methods for extracting the spiculation feature of a detected breast lesion on mammography by segmenting the contour of the lesion in a number of re- gions which are separately analysed, determining a characterizing spiculation feature set. In order to differentiate between benign and malignant tumors based on the extracted spiculation sets, an intelligent neural network is first trained on a number of 96 cases of known breast cancer malignancy and then tested for diagnosing and classifying breast cancer tumors. The input of the neural network is thus the ex- tracted spiculation feature set and the output is represented by the histopatological diagnostic given by doctors. Finally, the performance of the introduced methods is analysed depending on the number of regions in which the contour is segmented and the performance-related conclusions are stated for each of the methods. The highlight of this paper is the division of the tumour contour in regions and the assessment of a spiculation indicator for each region, resulting a set of spiculation indicators that characterise the tumour and - by training a neural network - can be used in classifying breast tumours with high performance. Keywords: breast cancer, spiculation feature extraction, neural network, diagnosis. 1 Introduction In the detection process of suspicious breast cancer lesions, mammography screening is usu- ally followed by an analysis of diagnostic mammography, i.e. a more detailed analysis of the artifacts visible on the mammographic imaging results, such as: breast tissue density, tumoral masses, (macro- and micro-) calcifications, architectural distortion, etc, [1]. Among them, the detected microcalcifications and the identified masses (also known as tumours) are considered the most relevant signs of malignancy that need to be investigated in detail. On one hand, the microcalcifications are tiny mineral deposits within the breast, representing an early indicator of the presence of a tumor. On the other hand, the tumoral mass is a tissue composed of an abnormal growth of cells (normal or neoplastic, cancerous, cells) with or without integrated calcifications that are also being detected on the mammographic results. When analyzing any pathology of breast cancer, the location, the size, the shape of the mass are usually assessed and the mass density and the tumoral margins are moreover evaluated. While the benign masses have usually smooth edges, are well circumscribed, compact and approximately circular or elliptical (see Figure 1), malignant lesions usually have vague edges and irregular form, presenting spicu- lations (i.e. a radial pattern of spiculs), [2]. Thus, in analyzing the tumoral mass, in addition to the lesion size, the degree of spiculation of the tumor edge is a particularly relevant indicator of malignancy (differences visible in Figure 1). Copyright c⃝ 2006-2013 by CCC Publications Breast Cancer Diagnosis based on Spiculation Feature and Neural Network Techniques 355 Figure 1: Benign and malignant tumors extracted from available mammographies The evaluation of the spiculations of masses visible on the mammographic results implies form recognition techniques that extract numerical information usable in the process of differentiating between benign and malignant tumours. These techniques are continuously being developed and improved in order to make possible a more refined characterization of the particular morfological features and a higher classification performance. [3] computes a spiculation index that can classify tumours with 93% accuracy; [4] proposes three methods for morfological feature extraction and reaches 93% accuracy when classifying breast lesions; [5] explores the classification power of 6 contour algorithms on 349 masses using three popular classifiers (Bayesian Classifier, Fisher Linear Discriminant Analysis and Support Vector Machine) and shows that a big variation (14%) of the quality of the segmentation method influences only 4% of the classification performance given by the contour features. Besides the current trend in technology and computer automation, the bioengineering inter- disciplinary developments of the last years allowed the implementation of digital CAD (Computer Aided Diagnosis) aid tools to assist novice radiologists in making diagnostic and recommended treatments decisions, herewith providing a second opinion on the decision and playing the role of the second pairs of eyes in the analysis, certifying or not the quality and/or the choice. Although the computational CAD techniques used in determining a medical prognostic do not always pro- vide the desired quality, their implementation using intelligent techniques (neural networks, fuzzy techniques, genetic algorithms and their hybrid variants) that are based on human experience and provide learning and adaptation abilities, demonstrates high performance in these kinds of tasks. In this paper, we are trying to define new methods for extracting the spiculation feature set of a breast lesion identified on a mammographic result by analyzing the lesion contour on a number of contour regions. The methods have been tested by using a neural architecture integrated into a CAD that provides good classification performance when trained on the extracted spiculation feature sets of 96 breast cancer cases with known malignancy diagnostic. Thus, this neural CAD is able to distinguish, diagnose and classify breast cancer tumors in benign or malignant, highlighting the benefit and usefulness of both the methods and their neural CAD application in national breast cancer screening programs. 2 Materials and methods In terms of imaging, the starry distortions visible on mammography are caused by the in- trusion of cancer in surrounding tissue (invasion occurs in malignant tumors), so that the con- tour/outline form of the tumor is generally correlated to the degree of malignancy. The contour is thus an extremely valuable information in the differentiation between benign and malignant tumors. As stated above benign tumors usually have smooth, circumscribed, macrolobulate and well defined contours (Figure 1), while malignant tumors are characterized by vague, irregular, 356 V. Bălănică, I. Dumitrache, L. Preziosi microlobulate and spiculate contours (Figure 1). Based on these observations, researchers have defined certain objective and quantitative measures/indicators to characterize the contours in order to improve the imaging classification capability of the detected tumors. In terms of perfor- mance, among the most important and most frequent contour indicators found in the literature we count: the degree of compactness, the spiculation index, the fractional concavity, the Fourier factor and the fractal dimension, [6]. In our paper, the neural networks are used for classifying the tumours based on the extracted spiculation feature set of a breast lesion identified on a mammographic result. The neural networks (NNs) models are simplified structural and functional nervous systems, formed by a number of processing elements, the neurons, which are bound by weighted connections, similar to the neural synapses, [7]. The neural network architecture mainly used in the medical field is known as the multi-layer perceptron (or MLP), [8], Figure 2, which consists of neurons usually organized in three feed-forward interconnected layers (i.e. all neurons in one layer are fully connected to all neurons in the next layer, but there are no feedbacks to previous layers). A perceptron has the ability to learn, to self-organize and to generalize (i.e. it can have same output for similar sets of inputs). The most commonly algorithm used for training the neural network is the back-propagation algorithm, which calculates the error gradient of the network and adjusts each connection weight by minimizing the mean square difference and achieving convergence to a local minimum, [9]. Figure 2 shows a graphical representation of such a network, where, in the case of a medical system, the input layer neurons often corresponds to the clinical symptoms, the hidden layer simulates the medical inference process done by the doctor and the neurons from the output layer correspond to the medical diagnosis. Figure 2: Graphical representation of a MPL neural network with one hidden layer The focus in this paper is not to develop a novel MPL, but rather to apply it in the clinical relevant problem of breast cancer. Because many clinical scenarios can not be explained based on only one parameter, but as interaction of various clinical-pathological factors, [10], the inputs of the MPL may include any set of mammographic descriptors of the calcifications (distribution, number, description), of the detected masses (size, shape, density, margins, location) or of any other associated findings (asymmetries, distorsions), being able to include also inputs related to the patient history. The ouputs of the MPL are usually numbers between zero and one that may correpond to the prediction of the biopsy outcome (benign or malignant) as in [11], the stage of development for the analyzed lesion (in situ or invasive) or it may be correlated to the survival chance/rate of the patient (under 1 year, under 2 years, under 3 years, under 5 years, more than 5 years) as in [12]. Breast Cancer Diagnosis based on Spiculation Feature and Neural Network Techniques 357 3 New methods for lesion spiculation feature extraction The mammographic lesions segmentation methods proposed below assumes that the spicules are visible on the mammography in the form of tumor infiltration or branches, usually narrow, in normal biological tissue. Therefore, the analysis of tumor contour on neighborhoods (or regions) of the contour and the measuring of the curvature change in each region offers an useful set of valuable information in determining the degree of spiculation in the analysed regions of the lesion contour and allowing subsequent CAD classification of tumors in benign or malignant. For each of the neighborhoods of the lesion contour four spiculation indicators are computed, namely: A. A Maximum Level Difference for every region of the contour, relative to the center of gravity of the lesion, according to the following figure (Figure 3): Figure 3: Calculation of the maximum level difference of for the regions of the tumoral contour B. The Total Area of triangles in a neighborhood, calculated by adding the triangle areas formed by each three consecutive points along the contour, according to the following figure (Figure 4): Figure 4: Calculation of the total area of the triangles formed along the contour for a contour region C. The Total Angle in a neighborhood, calculated by adding all the angles formed by every three consecutive points along the contour, according to the following figure (Figure 5): 358 V. Bălănică, I. Dumitrache, L. Preziosi Figure 5: Calculation of total angle summed for the angles formed along a contour region D. The Total Quadratic Curvature for each region, calculated by adding the curvatures of the circles passing through each three consecutive points along the contour, according to the following figure (Figure 6): Figure 6: Calculation of the total quadratic curvature, by adding the curvatures of the circles passing through each three consecutive points of the analysed region of the contour The main steps for the defined A, B, C and D methods are described as following: i. The starting point of the method is an image of a detected mass (Figure 7 a), ii. Calculate and plot the center of the mass based on the pixels that form the lesion (com- puting center of mass for lines and for columns) (Figure 7 b), iii. Calculate and plot the marginal pixels of the tumor (Figure 7 b), iv. Calculate and plot the distances between the center of mass and the pixels describing the edge of the tumor (Figure 7 c), v. The distance vector is ascendently ordered according to the angle formed by these distances lines with one reference line, vi. Depending on a relevant, experimentally determined, number of neighborhoods, X, for each angle Z = 360/X are done a number of operations specific for each algorithm. For each angle Z, for algorithm: Breast Cancer Diagnosis based on Spiculation Feature and Neural Network Techniques 359 A. Is computed the maximum difference of the distances (= center of gravity - marginal pixels) that fall within the Z angle, thus identifying the characteristic vector of spic- ulation for the tumor. Table 1 shows an example of this vector for a malignant and a benign lesion). B. Is computed the total area of triangles in the neighborhood described by the Z angle, by adding the areas of the triangles formed by every three consecutive points along the contour, according to the formula (1), [13]. Table 1 shows an example of this vector for a malignant and a benign lesion). ∑ iϵZ Ai, Ai = 1 2 [(Xi − Xi+2)(Yi+1 − Yi) − (Xi − Xi+)(Yi+2 − Yi)] (1) C. Is computed the total angle for the neighborhood described by the Z angle, by sum- ming of all angles formed by every three consecutive points along the contour, accord- ing to the formula (2), [13]. Table 2 shows an example of this vector for a malignant and a benign lesion). ∑ iϵZ Ui, Ui = arcsin[2·Ai/ √ (Xi − Xi+1)2 + (Yi − Yi+1)· √ (Xi+1 − Xi+2)2 + (Yi+1 − Yi+2)] (2) D. Is computed the total quadratic curvature for the neighborhood described by the Z angle, by adding quadratic curvatures of the circles passing through each three consecutive points along the contour, according to the formulas (3), (4), (5), [13]. Table 2 shows an example of this vector for a malignant and a benign lesion). ∑ iϵZ ci, ci = 1 R2i , Ri = √ d2 + e2 4a2 − f a , where (3) a = ∣∣∣∣∣∣∣ Xi Yi 1 Xi+1 Yi+1 1 Xi+2 Yi+2 1 ∣∣∣∣∣∣∣ , d = ∣∣∣∣∣∣∣ X2i + Y 2 i Yi 1 X2i+1 + Y 2 i+1 Yi+1 1 X2i+2 + Y 2 i+2 Yi+2 1 ∣∣∣∣∣∣∣ , (4) e = ∣∣∣∣∣∣∣ X2i + Y 2 i Xi 1 X2i+1 + Y 2 i+1 Xi+1 1 X2i+2 + Y 2 i+2 Xi+2 1 ∣∣∣∣∣∣∣ , f = ∣∣∣∣∣∣∣ X2i + Y 2 i Xi Yi X2i+1 + Y 2 i+1 Xi+1 Yi+1 X2i+2 + Y 2 i+2 Xi+2 Yi+2 ∣∣∣∣∣∣∣ (5) The spiculation features set computed with any of the four methods shows, after tests, a very good performance in terms of differentiating benign and malignant tumors and can be used to compute an objective measure/degree of spiculation for the examined lesions. Moreover, these methods are suitable for training a neural classifier of malignant or benign patterns of spicules that is able to make assessments in the presence of a new set of extracted mammographic spiculation features. As observed, depending on the analyzed lesion, the nature of these sets of indicators show a large variation of values in each neighborhood which makes them very suitable to being analyzed with classification modules such as neural networks in order to determine patterns and nonlinear correlations. The following practical approach shows the great classification potential of these indicators, given the fact that the trained neural modules reached maximum performance after a very short training time. 360 V. Bălănică, I. Dumitrache, L. Preziosi Figure 7: Lesion segmentation algorithm for computing the spiculation degree: a) a tumor identified on mammography, b) identification of the center of the mass and of the marginal pixels, c) calculating and sorting the angular distances from the center of the mass to the marginal pixels Table 1: The characteristical spiculation vectors for a malignant lesion and for a benign one for method A and method B However, the efficient utilization of these methods may vary depending on the size and the quality of the imaging results (processed or unprocessed, noise filtered or unfiltered regions of interest, the quality of image binarization, etc.) and in what concerns the choice of the number of neighborhoods for the segmentation of the lesion contour. In this respect, one can successfully use a genetic algorithm to determine the optimal number of neighborhoods to achieve the maximum classification performance. 4 Neural networks trained for breast cancer diagnosis The data sets obtained by applying the above methods can be used to train a neural intel- ligent module for the classification of the detected lesions on imaging results and determine the nonlinear correlations and the useful predictions. Thus, a feedforward neural network with one hidden layer, configurable in terms of the learning rate, of learning momentum, of the training epochs and of the minimum desired error, was trained using the backpropagation learning algo- rithm based on the extracted imaging spiculation feature sets determined with the above defined A, B, C and D methods (the feature sets will be noted as SetA, SetB, SetC and SetD). The Breast Cancer Diagnosis based on Spiculation Feature and Neural Network Techniques 361 Table 2: The characteristical spiculation vectors for a malignant lesion and for a benign one for method C and method D general training and testing algorithm of the neural module is given below: i. Having a selection of mammographies for which the diagnostic output is known, i.e. the diagnosis (benign or malignant), ii. The spiculation set for each of the mammographies is extracted with one of the A, B, C, or D methods described above, for a number of neighborhoods, iii. Both the input (the spiculation data set) and also the output (the diagnostic) are normal- ized, iv. Based on the normalized data, the neural network is being trained and validated for the predicton of the output (the diagnostic), v. The predicted output is compared with the actual output (truth values of malignancy) by calculating performance measures like: specificity (i.e. number of true negatives divided by the sum of true negatives and false positives), sensitivity (i.e. number of true positives divided by the sum of true positives and false negatives), accuracy (i.e. sum of true positives and true negatives divided by sum of true positives, false positives, true negatives and false negatives) and the Matthews Correlation Coefficient (MCC explained below). The training, the testing and the validation of the neural modules are based on the informa- tion of 96 mammographic cases acquired from the Department of Medical Physics, University of FreeState, Bloemfontain, South Africa, of which 48 cases are benign and 48 are malignant, [14]. The practical experience has focused on the analysis of several selections of features in order to determine the training and classification performances of the neural networks. Rapid convergence of the training error and the classification and prediction power of the neural network indicate the correlation strength/weight between inputs and outputs, directly justifying the existence or the absence of non-linear correlations between them, in this case evaluated by Matthews Correlation Coefficient (MCC) - belonging to the interval [-1, +1], +1 representing a perfect correlation, 0 an arbitrary correlation and -1 inverse correlation. 362 V. Bălănică, I. Dumitrache, L. Preziosi Because the quality of the spiculation characteristics introduced in the previous section may vary depending on the number of neighborhoods set for segmentation of the lesions contour, the tests include also the variation of this value. Simultaneously, the regions of interest (ROI) processed and used in evaluating the imaging characteristics have the dimension of 335x436 pixels, an increase of the size being possible to the expense of the time needed for extracting data and for running the tests. On average, 5 to 8 neural networks were trained for the same set of features in order to observe and determine the recorded performance variation. The performed tests (see Table 3) are not exhaustive, the results being obtained based on the currently available data sets. 5 Discussion of results The performance results presented in Table 3 give a clear indication about the high potential of the introduced spiculation-feature extraction methods in what concerns the differentiation between benign and malignant tumors when using a neural classifier. However, as shown in Table 4, the developed methods differ in terms of power and speed of convergence of the classifications and provide a different classification stability when changing the number of analyzed lesion contour neighborhoods. Thus, the method that calculates the maximum level difference on each region of the contour offers the optimal method for extracting the spiculation feature and the algorithm that calculates the total quadratic curvature for each region offers the most reliable measure of the lesion malignancy regardless of the neighborhoods number. 6 Conclusions This paper introduces four new methods for the assessment of the tumor contour on a number of boundary neighborhoods. The Maximum Level Difference is calculated for every region of the analyzed contour, relative to the center of gravity of the lesion. The Total Area of Triangles in a neighborhood is calculated by adding the areas of each triangle formed by three consecutive points along the contour. The Total Angle for a neighborhood is calculated by adding up all the angles formed by every three consecutive points along the contour. The Total Quadratic Curvature for each region is calculated by adding the curvatures of the circles passing through each three consecutive points along the contour. The spiculation feature extraction algoritms are tested on a selection of mammographies with known diagnostic by using a neural network that is trained on the extracted feature sets. Following the performance tests, the spiculation feature sets computed with all of the four methods show a high potential in the differentiation of benign and malignant tumors and can be used to compute an objective measure/degree of spicularity for the analysed lesion. Moreover, these methods are suitable for training a neural classifier of malignant or benign patterns of spiculs that is able to make assessments in the presence of a new set of extracted mammographic spiculation features. However, as shown in the pages of the paper, the developed methods differ in terms of power and speed of convergence of the classifications and provide a different classi- fication stability when changing the number of neighborhoods the lesion contour is segmented and analyzed. These CAD extraction methods show high reliability and prove high implementation potential being currently used in an automated decision support systems for breast cancer designed for national mammography screening programs in Romania and South Africa. Breast Cancer Diagnosis based on Spiculation Feature and Neural Network Techniques 363 ID No. of neighbor- hoods Inputs of the neural network Outputs of the neural network Malignity classification per- formances MCC Correla- tion 1 5 SetA =5 inputs Diagnostic Sensitivity: [ 97 100 %] Specificity: [ 100 98.1 %] Accuracy: [ 98.9 %] For 50-200 training iterations 0.97 2 5 SetB =5 inputs Diagnostic Sensitivity: [ 86.3 100 %] Specificity: [ 96.2 83 %] Accuracy: [ 91.7 90.7 %] For 100-300 training iterations 0.83 3 5 SetC =5 inputs Diagnostic Sensitivity: [ 90.9 93.1 %] Specificity: [ 90.5 98.1 %] Accuracy: [ 90.7 95.8 %] For 100-300 training iterations 0.86 4 5 SetD =5 inputs Diagnostic Sensitivity: [97.7 100 %] Specificity: [ 98.1 98.1 %] Accuracy: [ 97.9 98.9 %] For 50-100 training iterations 0.98 5 5 The best sets: SetA,SetB =10 inputs Diagnostic Sensitivity: [ 100 %] Specificity: [ 100 %] Accuracy: [ 100 %] For 20-40 training iterations 1 6 5 SetA,SetB, SetC,SetD =20 inputs Diagnostic Sensitivity: [87.7 91.6 %] Specificity: [ 97.9 100 %] Accuracy: [ 92.7 95.8 %] For 100-300 training iterations 0.90 7 20 SetA =20 inputs Diagnostic Sensitivity: [ 100 %] Specificity: [ 100 %] Accuracy: [ 100 %] For 20-30 training iterations 1 8 20 SetB =20 inputs Diagnostic Sensitivity: [ 100 %] Specificity: [ 100 %] Accuracy: [ 100 %] For 50-150 training iterations 1 9 20 SetC =20 inputs Diagnostic Sensitivity: [ 100 %] Specificity: [ 100 %] Accuracy: [ 100 %] For 4-10 training iterations 1 10 20 SetD =20 inputs Diagnostic Sensitivity: [ 100 %] Specificity: [ 100 %] Accuracy: [ 100 %] For 150-200 training iterations 1 11 20 The best sets: SetA,SetC =40 inputs Diagnostic Sensitivity: [ 100 %] Specificity: [ 100 %] Accuracy: [ 100 %] For 3 training iterations 1 12 20 SetA,SetB, SetC,SetD =80 inputs Diagnostic Sensitivity: [ 100 %] Specificity: [ 100 %] Accuracy: [ 100 %] For 3-4 training iterations 1 Table 3: The trained neural networks and the achieved performances 364 V. Bălănică, I. Dumitrache, L. Preziosi Set ID Description of the Spicula- tions Feature Set Observed features Description SetA Spiculation set for the A algorithm: The Maximum Level Difference for every region of the contour Average con- vergence and avarage stabil- ity Regardless of the neighbor- hoods number, the classifi- cation performance converges with average speed, and the obtain results vary very little. SetB Spiculation set for the B algorithm: The Total Area of Triangles in a neighborhood Slow conver- gence and low stability Regardless of the neighbor- hoods number, the classifi- cation performance converges with low speed. For a small number of regions, the re- sults are much weaker com- pared with those obtained for a higher number of regions. SetC Spiculation set for the C algorithm: The Total Angle in a neighborhood Fast conver- gence and low stability Regardless of the number of neighborhood, the classifica- tion performance converges with high speed. For a small number of regions, the results are poor compared to those obtained for a larger number of regions. SetD Spiculation set for the D algorithm: The Total Quadratic Curvature for a region Slow conver- gence but very high stability Regardless of neighborhoods number, the classification per- formance converges with low speed, but the results are al- ways powerful and do not vary. Table 4: The efficiency of the spiculation feature extraction methods for a lesion Breast Cancer Diagnosis based on Spiculation Feature and Neural Network Techniques 365 Bibliography [1] Sultana, Alina (2010); On Improving Image-Based Diagnosis Using Digital Image Processing, Faculty of Electronics, Telecommunications and Information Technology, Bucharest. [2] Feig, S.A.; Yaffe, M.J. (1995); Digital Mammography, Computer-Aided Diagnosis and Tele- mammography, The Radiologic Clinics of North America, Breast Imaging Press, 33(6):1205- 1230. [3] Guliato, D.; Rangayyan, R.M.; Carvalho, J.D.; Santiago, S.A. (2006); Spiculation-preserving Polygonal Modeling of Contours of Breast Tumors, Proceedings of the 28th IEEE, 2791-2794. [4] Cheikhrouhou, I.; Djemal, K.; Sellami, D.; Maaref, H.; Derbel, N. (2008); New mass descrip- tion in mammographies, Image Processing Theory, Tools and Applications. [5] DomĂnguez, A.R.; Nandi, A.K. (2009); Toward breast cancer diagnosis based on automated segmentation of masses in mammograms, Pattern Recognition, 42(6):1138-1148. [6] Rangayyan, R.M.; Nguyen, T.M. (2005); Pattern classification of breast masses via fractal analysis of their conturs, International Congress Series, 1281:1041-1046. [7] Dumitrache, I.; Buiu, C. (1995); Hybrid geno-fuzzy controllers, IEEE Intelligent Systems for the 21st Century, 5:2034-2039. [8] Pandey, B; Mishra, R.B. (2009); Knowledge and intelligent computing system in medicine, Computers in Biology and Medicine, 39:215-230. [9] Alpaydin, Ethem (2010); Introduction to machine learning, MIT Press, 2010. [10] Drew, P. J.; Monson, J. R.T. (2000); Artificial neural networks, Surgery, 127:3-11. [11] Marcano-Cedeno, A.; Quintanilla-Dominguez, J.; Andina, D. (2011); Breast cancer classifi- cation applying artificial metaplasticity algorithm, Neurocomputing, 74:1243-1250. [12] Mofidi, R.; Deans, C.; Duff, M.D.; Beaux, A.C.; Brown, S. P. (2006); Prediction of survival from carcinoma of oesophagus and oesophago-gastric junction following surgical resection using an artificial neural network, European Journal of Surgical Oncology, 533-539. [13] Hazewinkel, M. (1994); Encyclopaedia of Mathematics (set), Kluwer, ISBN 1-55608-010-7. [14] Balanica, V.; Rae, W.I.D.; Caramihai, M.; Acho, S.; Herbst, C.P. (2009); Integration of image and patient data, software and international coding systems for use in a mammography research project, World Academy of Science, Engineering and Technology, 58:1002-1005.