Accurate classification of breast tissue types in mammograms is crucial for effective diagnosis, but variations in breast density and imaging noise present significant challenges. This study investigates the robustness of Haralick texture analysis for classifying breast tissues into 4 categories: fatty, glandular and connective, dense, and abnormal, to enhance classification accuracy. Three distinct strategies were evaluated: first, extracting Haralick features directly from the selected regions of interest (ROIs) without preprocessing; second, applying various preprocessing techniques globally to the entire mammogram, then cropping the ROI; and third, using the same preprocessing methods locally to the cropped ROIs. For each strategy, 148 mammograms from the Mini-MIAS database were analyzed. ROIs were cropped into 40×40-pixel subimages. Texture features were extracted using Gray-Level Co-occurrence Matrix (GLCM)-based Haralick descriptors within a 20×20 sliding window, yielding a total of 592 samples for analysis. Linear Discriminant Analysis (LDA) was employed for classification, and model performance was evaluated using various metrics. The strategy of extracting features without prior preprocessing achieved the highest classification accuracy of 98.5%, with strong sensitivity and specificity across all classes. This study highlights the potential of Haralick texture descriptors as a reliable tool for early detection of breast cancer.
Ali et al. (Wed,) studied this question.