Scholarly article on topic 'Mammogram Classification using Law's Texture Energy Measure and Neural Networks'

Mammogram Classification using Law's Texture Energy Measure and Neural Networks Academic research paper on "Computer and information sciences"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Computer Science
OECD Field of science
Keywords
{Mammography / LAWS / GLCM / "Texture ;"}

Abstract of research paper on Computer and information sciences, author of scientific article — Arden Sagiterry Setiawan, Elysia, Julian Wesley, Yudy Purnama

Abstract Mammography is the best approach in early detection of breast cancer. In mammography classification, accuracy is determined by feature extraction methods and classifier. In this study, we propose a mammogram classification using Law's Texture Energy Measure (LAWS) as texture feature extraction method. Artificial Neural Network (ANN) is used as classifier for normal- abnormal and benign-malignant images. Training data for the mammogram classification model is retrieved from MIAS database. Result shows that LAWS provides better accuracy than other similar method such as GLCM. LAWS provide93.90% accuracy for normal-abnormal and 83.30% for benign-malignant classification, while GLCM only provides 72.20% accuracy for normal-abnormal and 53.06% for benign-malignant classification.

Academic research paper on topic "Mammogram Classification using Law's Texture Energy Measure and Neural Networks"

(8)

CrossMark

Available online at www.sciencedirect.com

ScienceDirect

Procedía Computer Science 59 (2015) 92 - 97

International Conference on Computer Science and Computational Intelligence (ICCSCI 2015)

Mammogram Classification using Law's Texture Energy Measure and Neural Networks

Arden Sagiterry Setiawan1, Elysia1*, Julian Wesley1 and Yudy Purnama1

!Master of Information Technology, Bina Nusantara University

Jl KH Syahdan 9, Jakarta 11480, Indonesia

Abstract

Mammography is the best approach in early detection of breast cancer. In mammography classification, accuracy is determined by feature extraction methods and classifier. In this study, we propose a mammogram classification using Law's Texture Energy Measure (LAWS) as texture feature extraction method. Artificial Neural Network (ANN) is used as classifier for normalabnormal and benign-malignant images. Training data for the mammogram classification model is retrieved from MIAS database. Result shows that LAWS provides better accuracy than other similar method such as GLCM. LAWS provide93.90% accuracy for normal-abnormal and 83.30% for benign-malignant classification, while GLCM only provides 72.20% accuracy for normal-abnormal and 53.06% for benign-malignant classification.

© 2015TheAuthors. Published by ElsevierB.V.This is an open access article under the CC BY-NC-ND license (http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-reviewunderresponsibilityoforganizingcommittee of the International Conference on Computer Science and Computational Intelligence (ICCSCI 2015)

Keywords:Mammography; LAWS; GLCM; Texture;

1. Introduction

Breast cancer is the most common type of cancer which can be found among women. The abnormality can be found hiding below the breast tissue structure. Actually, there are several ways to detect early indication of breast cancer such as mammography, biopsy, ultra sound image, and thermography1. Among these methods, mammography is the best approach in early detection of breast cancer. However, mammography method is

* Corresponding author. Tel.: +62-21-5345830 E-mail address: elysia@binus.edu

1877-0509 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of organizing committee of the International Conference on Computer Science and Computational Intelligence (ICCSCI 2015) doi: 10.1016/j.procs.2015.07.341

challenging. Mostly the given visual clues may be subtle. The visual representation can also varies even for images in the same categories.

Feature extraction is the first step in digital mammogram process. In mammography classification, accuracy is determined by feature extraction methods and classifier. Texture feature is very important in image processing2. The aim is to segments several images based on their similarity in texture.

There are several methods for processing texture features. Earlier studies was using texture analysis from Spatial grey level dependencee (SGLD) matrices to predict whether the presence of microcalcifications is associated with malignant or benign pathology. Artificial Neural Network classifier was used to test the texture in classifying malignant or benign. The result was 11 of the 28 benign cases were correctly identified (39 % specificity) without missing any malignant cases (100 % sensitivity)3.

Another method is Gray Level Co-occurrence Matrix (GLCM). GLCM calculate the probability of two pixels having particular gray level at particular spatial relationship. This method has been widely used for texture feature extraction4.

There is also Laws' texture energy measures (LAWS) developed by K. I. Laws whichhas been used for variant applications. This method combines predetermined one-dimensional kernels into various convolution masks5. LAWS potentially is superior to GLCM because it extracts second layer of the image and it is more accurate. Some experimental results show that Laws' method has high capability in feature texture extraction that can detect cancer5.

In this study, we build a mammogram classification using LAWS as feature extraction method and Artificial Neural Network (ANN) as classifier. ANN isa nonlinear model for forecasting and included in data-driven self-adaptive methods6. ANN can learn from examples and captures the functional relationships among the hard description of data.

2. Method

fig 1. Classification Methodology

Data Collection

Data Source for this research is taken from The Mammographic Image Analysis Society (MIAS). MIAS is an organization of UK research groups interested in the understanding of mammograms and has generated a database of digital mammograms. There are 327 images available for download at http://peipa.essex.ac.uk/pix/mias/. The database has been reduced to 200 micron pixel edge. Therefore all images are sized at 1024x1024 pixels.

Data Pre-processing

Matlab is used for this whole experiment.Original size of the images is 1024x1024 pixels. We decide to crop the image into smaller size to increase processing speed. We crop the image to fit their region of interest. The x and y image-coordinates of centre is given in the image database. We decide best fit size is 128x128 pixels. Figure 2 show images identified as normal condition while figure 3 show images identified as abnormal. We only crop the region based on its region of interest.

fig 2. Pre-processing for normal image

fig 3. Pre-processing for abnormal image

LTEM Based Feature Extraction

Laws' texture energy measures (LAWS) is a method to extract secondary features of the image, LAWS uses filters mask within fixed window size7. LAWS is chosen because of its superior capability to extract texture of an image.There are many features which can be extracted from an image. This could be colour, shape or texture. Texture is important feature for classification purpose. Texture is extracted to attain higher information of an image. Texture can be qualitatively described by its coarseness. The value might be repetitive under same local region and this can become strong identifier among similar images.

LAWS uses a set of 5x5 convolution masks is used to compute the texture energy. This texture energy then represented as a vector of nine numbers for each pixel of the image being analysed. There are 4 main characteristics of the image can be analysed: level, edge, spot and ripple.

L5 (level) = [1 4 6 4 1]

E5 (level) = [-1 2 0 2 1]

S5 (spot) = [-1 0 2 0-1]

R5 (ripple) = [1 -4 6 -4 1]

Each capital letter in front of the vector name indicates the characteristic it represents. The L5 vector gives the calculation of centre-weighted local average E5 vector indicates edges, S5 vector detect spot and R5 vector detect ripple. Laws procedure begins with moving small window around the images. LAWS subtracts the local average from each pixel to produce new processed image, in which the average intensity of each neighbourhood is near to zero. In this study we use 15x15 pixels window. We can produce 2D convolution masks by computing outer product of each individual vector.

Table 1. Combination of Each Vector

L5 E5 S5 R5

L5 L5L5 E5L5 S5L5 R5L5

E5 L5E5 E5E5 S5E5 R5E5

S5 L5S5 E5S5 S5S5 R5S5

R5 L5R5 E5R5 S5R5 R5R5

Table 1 shows the 16 combined vectors. L5E5 measures vertical edge content while E5L5 measures horizontal edge content. The average of those two will become total edge content. We do not use L5L5 because the sum of the filter elements is not zero. In the end there will be 9 vectors, which are shown in table 2:

Table 2. Chosen Vectors

Chosen Vectors

L5E5/E5L5 R5S5/S5R5

L5S5/S5L5 S5S5

L5R5/R5L5 E5E5

E5S5/S5E5 R5R5 E5R5/R5E5

The final result will be 9 energy images for each single image. Actually it is a single image with 9 texture attributes at each pixel. We calculate the statistical values of each image. We decide to calculate their mean (equation 1) and variance (equation 2). W is the size of the window. These numbers then will be used as input for the next classification process.

neighbouring pixels

mean =-—--(1)

variance =

„{neighbouring pixels — mean)2 W

For the comparison, we prepare GLCM as feature extraction method. There are 16 (sixteen) texture features extracted using GLCM. These texture features was selected into four (4) most discriminant features using a statistical multi-variate t-test as suggested by previous research [8]. The texture features are ASM (energy), correlation, sum of variance and difference entropy.

ANN Classifier

ANN is used as classifier. We choose ANNbecause of its capabilityto learn from examples and capture the functional relationships among the hard description of data. Figure 4 show our ANN classifier model.

Input Node Hidden Node Output Node

fig 4. ANN Model, both Normal-Abnormal and Benign Malignant

This ANN is two-layer feed-forward back- propagation network with sigmoid transfer functions in both the hidden layer and the output layer. The back-propagation is based on levenberg-marquardt. Composition of the data is: 70% training data, 15% validation data, and 15% testing data.There will be 18 input nodes. We train the network several times with different amount of hidden layer each training process. Best result attain at 10 hidden nodes. The output node is taken from notes given from MIAS database.

There will be two steps of classification. First is classification of normal and abnormal data. Second one, the abnormal data gathered from first classification will be classified again to determine whether they are benign or malignant.

For comparison purpose, we have also prepared another similar ANN design. However this time we use GLCM as feature extraction method. The only difference with previous ANN is design is the hidden node for this classifier is 4 nodes.

3. Results and Discussion

The neural network model based on LAWS features is applied to classify mammogram images from MIAS database. Two neural network models were created, one for classifying normal and abnormal class, and another one for classifying benign and malignant class. The result of classification for LAWS features is compared with classification result of neural network model based on GLCM features (table 3).

Table 3. Classification accuracy

LAWS GLCM GLCM GLCM GLCM

0° 45° 90° 135°

NormalAbnormal Specificity Sensitivity Accuracy 1.00 0.91 93.90% 0.50 0.65 63.30% 1.00 0.72 73.50% 0.25 0.76 71.40% 0.33 0.65 63.30%

Benign-Malignant Specificity Sensitivity Accuracy 0.88 0.80 83.30% 0.62 1.00 66.70% 0.50 0.75 72.20% 0.67 0.67 66.70% 0.50 0.67 66.10%

Neural network models classification based on LAWS features yield 93.90% accuracy for normal-abnormal classification and 83.30% for benign-malignant classification.Experiment on GLCM features is performed on four possible degrees of orientation: 0-degrees, 45-degrees, 90-degrees, and 135-degrees. For normal-abnormal classification, 0-degrees model yield 63.30% accuracy, 45-degrees model yield 73.50% accuracy, 90-degrees model yield 71.40% accuracy, and 135-degrees model yield 63.30% accuracy. For benign-malignant classification, 0-degrees model yield 66.70% accuracy, 45-degrees model yield 72.20% accuracy, 90-degrees model yield 66.70% accuracy, and 135-degrees model yield 66.10% accuracy.

Calculation of accuracy for normal-abnormal class and benign-malignant class was performed separately. For benign-malignant class, the training data was all of abnormal data class from the entire data set. As a result, calculation of accuracy for benign-malignant class should be done with consideration of normal-abnormal class accuracy.

For LAWS features, the true accuracy value of benign-malignant classification would be 78.21%. GLCM features gives 42.22% accuracy for 0-degrees model, 53.06% accuracy for 45-degrees model, 47.62% accuracy for 90-degrees model, and 41.84% accuracy for 135-degrees model.

4. Conclusion

In this paper, we studied the usage of LAWS features as descriptors for classifying mammogram images. Based on result of the experiment, LAWS features give better accuracy when classifying mammogram images compared to GLCM features. The true accuracy value of benign-malignant classification is 78.21%, but using GLCM feature, the accuracy less than 55% for each degrees.LAWS features also provide higher sensitivity and specificity compared to GLCM features.

In future works, improvement can be done by changing the architecture of neural network model. We can try to improve the accuracy of classification by changing the number of nodes in hidden layer. Another area to explore is by using different architecture for neural network model, such as radial basis function network.

5. References

[1] B. Santhi and R. Nithya, "Comparative Study on Feature Extraction Method for Breast Cancer Classification," Journal of Theoretical and Applied Information Technology , vol. 33, no. 2, pp. 220-226, 2011.

[2] A. Kandaswamy and H. Sheshadri, "Detection of breast cancer by mammogram image segmentation," J Cancer Res Ther, vol. 1, no. 4, pp.

232-234, 2005.

[3] H.-P. Chan, et al., "Computerized classification of malignant and benign micriocalcifications on mammograms: texture analysis using an artificial neural network," 1997.

[4] K. A. Mohd, R. Besar, Z. W. Wan, and N. Ahmad, "Identification of masses in digital mammogram using gray level co-occurrence matrices," Biomedical Imaging and Intervention Journal, pp. 1-13, 2009.

[5] H. A. Elnemr, "Statistical Analysis of Law's Mask Texture Features for Cancer and Water Lung Detection," vol. 10, no. 6, 2013.

[6] G. Zhang, B. E. Patuwo, and M. Y. Hu, "Forecasting with artificial neural networks: The state of the art," 1998.

[7] K. I. Laws, "Textured Image Segmentation," 1980.

[8] C.-H. Wei, C.-T. Li, and R. Wilson, A Content-Based Approach to Medical Image Database Retrieval. 2006.