Texture-based feature selection for Breast Ultrasound Classification Using Sum and Difference Histograms

Abstract

Automatic analysis of lesions in breast ultrasound images is a valuable support tool for the early diagnosis of breast cancer; however, speckle noise, low contrast, and variability in tissue appearance pose challenges for classification systems. This work proposes a methodology for classifying breast ultrasound images using Machine Learning algorithms, based on a feature descriptor derived from the Sum and Difference of Histograms (SDH). We conducted the experimental evaluation using the publicly available Breast Ultrasound Images (BUSI) dataset, which comprises 780 images distributed among 437 benign, 210 malignant, and 133 normal cases. To determine the most suitable configuration, we evaluated different preprocessing techniques, multiple window sizes, and 24 Machine Learning (ML) algorithms, and automatically selected the combination that achieved the best performance. The results show that the proposed descriptor adequately discriminates among breast lesion categories; the best configuration used a Boosted Trees classifier with a 21Å~21 window, achieving an F1-score of 90.35% and an accuracy of 86.15%. These results indicate that the proposed methodology offers a lower-computational-complexity alternative to Deep Learning (DL)-based approaches while maintaining competitive performance for developing computer-aided diagnosis systems.