Chapter 6 - Optimization of ROI Size for Development of Computer Assisted Framework for Breast Tissue Pattern Characterization Using Digitized Screen Film Mammograms

Abstract

In this work exhaustive experimentations have been performed for the selection of optimum region of interest (ROI) size for development of a computer assisted framework for breast tissue pattern characterization using digitized screen film mammograms. The classification of breast density or breast tissue pattern depends upon (a) the location and (b) the size of ROI. It has been well-established that the maximum information about the breast density is present at the heart of the breast where the glandular tissues are present in rich amount, thus in this study various experiments have been conducted to check the result of varying ROI size cropped from glandular ducts area for 4-class breast tissue pattern and 2-class breast tissue pattern characterization tasks using various texture feature models. These experiments have been conducted on digital database for screening mammography (DDSM) dataset mammograms. However, for 2-class breast tissue pattern characterization task, images belonging to {BIRADS-I, BIRADS-II} classes have been considered in “fatty” class and images belonging to {BIRADS-III, BIRADS-IV} classes have been considered in “dense” class. The work has been carried out on a total of 480 images, that is, 120 images from each image class (120×4=480). From each image, ROI of different sizes, that is, 32×32, 64×64, 128×128, and 256×256 pixels have been manually extracted from the core region of the breast tissue where glandular tissues are available in prominent amount. Various texture features have been extracted from each ROI using (a) statistical texture feature models, (b) signal processing based texture feature models, and (c) transform domain texture feature models. The computed texture feature vectors are inputted to the support vector machine (SVM) classifier. From the exhaustive experiments conducted in this study, it can be concluded that (a) out of the four different ROI sizes higher prediction rate have been observed using ROI size of 128×128 pixels, using all the texture feature models for 4-class breast tissue pattern characterization as well as for 2-class breast tissue pattern characterization, (b) the value of prediction rate of 79.5% has been attained for 4-class breast tissue pattern characterization using GLCM texture features with interpixel distance “d”=10 using SVM classifier, (c) the highest prediction rate value of 91.2% has been obtained for 2-class breast tissue pattern characterization using Laws’ masks of kernel width 3 with SVM classifier. The results obtained from the study conclude that the ROI size of 128×128 pixels manually extracted from the core region of the breast where glandular tissues are available in prominent amount that provides significant amount of clinical information for the differentiations between various breast tissue patterns.