Texture-Based Breast Cancer Nodule Classification in Ultrasound Imaging Using Adaptive Median Filtering and MLP

Abstract

The Ultrasonography (USG) is an imaging technique used to identify breast abnormalities.  The benefits of ultrasound imaging encompass its non-invasive nature and absence of radiation exposure.  Nonetheless, ultrasound imaging outcomes often involve issues that cause variations in physicians' interpretations of breast ultrasound images.  Computer-Aided Diagnosis (CAD) is a computational approach that offers an objective secondary assessment in identifying the attributes of nodules in breast ultrasound images.  The CAD method includes an image preparation phase that encompasses RoI selection, filtering, texture feature extraction, and classification.  The filtering procedure is executed utilizing a median adaptive filter and a median filter.  Texture feature extraction is performed with nine histogram features and 21 Gray Level Co-occurrence Matrix (GLCM) features.  Extraction results from Scilab indicate that employing 30 texture features enables the Multi Layer Perceptron (MLP) to categorize cystic and solid mass nodules with an accuracy of 88.89%, while utilizing 10 texture features yields an accuracy of 80.56%.