Tonal and Texture Features to Differentiate between Mass and Normal Breast Tissue on Digital Mammograms
Celia Varela1, Pablo G. Tahoces2, Arturo J. Méndez2, Miguel Souto1, Juan J. Vidal1
1Department of Radiology of the University of Santiago de Compostela, SPAIN
(Complejo Hospitalario Universitario de Santiago)
2Department of Electronics and Computer Science of the University of Santiago de Compostela, SPAIN
Introduction
Computer-aided diagnosis (CAD) schemes are being developed to provide a second opinion to the radiologists in order to improve detection rate of breast cancer1.
Morphological, tonal, and texture features have been used to decrease the number of false positive detections on CAD programs. Features extracted from regions of interest (ROIs) have been used in a variety of classifiers, based on either neural networks or discriminant analysis, to discriminate between mass and normal tissue1.
The aim of this study was to develop a series of tonal and texture features to differentiate between mass and normal breast tissue on digital mammograms. They were used as inputs into a backpropagation neural network (BPN) classifier to discriminate between normal and abnormal tissue.
Material and Methods
Sixty mammograms, each having a single biopsy proven malignant mass, were used. They were digitized at a resolution of 2000x2600 pixels and 1024 gray levels.
Three different ROIs were manually selected from each mammogram, one containing a mass, one with dense tissue, and one with fatty tissue. Thus, a total of 180 ROIs were used to extract tonal and texture features. The database was randomly divided into training (G1) and test (G2) groups.
Eight different tonal features were determined. Six features based on the absolute values of the gray levels of the pixels into the ROI and two features being developed based on the changes in the gray levels values of the pixels into the ROI when different resolutions were used. Moreover, two texture features were evaluated: coarseness and contrast as described by Amadasun and King2.
Correlation and variance-normalized distance were calculated to evaluate the ability of the features to differentiate between mass and normal tissue. These values provided a guide for selection of features to be introduced into the BPN classifier. Different feature combinations were introduced to the classifier to study the feasibility of the different features for the classification task.
A three-layered feed-forward BPN was developed to be used as a classifier. The delta-bar-delta learning rule was used during the training process3.
G1 was used to train and optimize the BPN while G2 was used to evaluate the classifier. The area under the ROC curve was used as index of classification accuracy of the BPN.
Results
A BPN was applied as a classifier on ROIs to discriminate between mass and normal breast tissue. We have obtained a best Az value of 0.92. Both, training and test ROC curves are shown in Fig. 1.
Fig. 1. ROC curves for training and test groups.
Discussion
The use of tonal and texture features is advantageous because they can be applied over regions without the need of accurately extracting the border of the lesions, as opposed to morphological or shaped-related features.
A set of tonal and texture features to differentiate between mass and normal breast tissue was developed. A BPN classifier was developed to distinguish between mass and normal tissue. The combination of using both types of features improves the performance of the classifier over using either type of feature alone. Our results establish the feasibility of using an artificial neural network classifier with tonal and texture features for classification of mass and normal tissue. This classifier could be implemented in a CAD scheme to reduce false positive detections.
References
1.- Wei D, Chan H-P, Petrick N, Sahiner B, Helvie MA, Adler DD, and Goodsitt MM. “False-positive reduction technique for detection of masses on digital mammograms: Global and local multiresolution texture analysis”. Med. Phys. 24(6), pp. 903-914 (1997).
2.- Amadasun M, King R. “Textural features correspoding to textural propierties”. IEEE Trans. on Systems, Man and Cybernetics 19, pp. 1264-1274, (1989).
3.- Jacobs RA. “Increased rates of convergence through learning rate adaptation”. Neural Networks 1, pp. 295-307, (1988).
Corresponding author:
Celia Varela
Departamento de Radiología
Facultad de Medicina
C/ San Francisco, 1
15704 Santiago de Compostela
Spain
Phone: 34 - 81- 570982
Fax: 34 - 81- 547031
e-mail: mrcuca(at)uscmail.usc.es
Oral presentation at EuroPACS'98, Barcelona, Spain