Study of Classification of Breast Lesions Using Texture GLCM Features Obtained From the Raw Ultrasound Signal

• Main Authors: Mariusz Nieniewski, Leszek J. Chmielewski
• Format: Article
• Language: English
• Published: Slovenian Society for Stereology and Quantitative Image Analysis 2020-06-01
• Series: Image Analysis and Stereology
• Subjects: breast lesion classification; quantitative ultrasound; feature selection; texture analysis; stepwise logistic regression
• Online Access: https://www.ias-iss.org/ojs/IAS/article/view/2113

Most of the methods of classification of breast lesions in ultrasound (US) images have been tested on B-mode images from the commercial equipment. The new possibility of further analysis of this problem showed up with the availability of a public database containing original raw radio frequency (RF) signals. In particular, it appeared that the original texture might contain diagnostic information which could be modified in the typical image processing and which is more difficult to perceive than the details of lesion shape/contour. In this paper a detailed analysis of the lesion texture is conducted by means of the decision trees and logistic regression. The decision trees turned out useful mainly for selecting texture features to be employed in the stepwise logistic regression. The RF signals database of 200 breast lesions was used for testing the performance of the benign vs malignant lesion classifier. The Gray Level Cooccurrence Matrix (GLCM) was calculated with the vertical/horizontal offset of up to five pixels. For each of these matrices six features were calculated resulting in a total of 210 features. Using these features a sufficient number of decision trees were generated to calculate pseudo-Receiver Operating Characteristics (ROCs). The outcome of this process is a collection of generated trees for which the employed features are known. These features were then used for generating generalized linear model by means of stepwise logistic regression. The analyzed regression models included the coefficients of up-to-the second degree terms. The texture features were further completed by a single shape feature, that is tumor circularity. The automatic procedure for finding the exact mask of a lesion is also provided for the conditions when the acoustic shadowing makes it impossible to obtain the entire contour reliably and a half-contour has to be used. The selected logistic regression models gave ROCs with the Area Under Curve (AUC) of up to 0.83 and the 95 % confidence region (0.63 0.96). Analyzing classification results one comes to the conclusion that the tumor circularity, which is the most informative among shape/contour features, is not essential against the background of textural features. The reported study shows that a relatively straightforward procedure can be employed to obtain benign vs malignant classifier comparable with that originally used for the database of the raw RF signals and based on the more complicated segmentation of the parameter maps of homodyned K distribution.