Discriminating Benign and Malignant Breast Lesions Using Shape Descriptors in Ultrasound Syahril Siregar1, Zahra Azizah2, Djarwani Soeharso Soejoko1
1. Department of Physics, Faculty of Mathematics and Natural Sciences, Universitas Indonesia, Depok 16424, Indonesia
2. Department of Computer and Informatics Engineering, Politeknik Negeri Jakarta, Jl. Prof. Dr. G.A. Siwabessy, Kampus Universitas Indonesia Depok, 16425, Indonesia
Abstract
Accurate differentiation between benign and malignant breast lesions in ultrasound imaging is essential for timely and appropriate clinical decision-making. While convolutional neural networks (CNNs) have demonstrated high classification performance, their lack of interpretability poses challenges in understanding the morphological basis of their predictions. To address this, the present study emphasizes the use of interpretable, shape-based descriptors rather than texture-based features. Six Shape Factor Analysis metrics-minimum-to-maximum axis ratio, circularity, solidity, extent, elongation, and eccentricity-were extracted from a publicly available dataset containing labeled ultrasound images of breast lesions. Partial Least Squares Discriminant Analysis (PLS-DA) applied to these metrics revealed a clear separation between benign and malignant categories using three latent variables, achieving a cross-validated accuracy of 0.924 and an F1-score of 0.857. All shape features contributed significantly to class distinction, with solidity, extent, and elongation identified as the most influential. Benign lesions exhibited higher values in circularity, solidity, extent, elongation, and eccentricity, reflecting their typically regular and well-defined morphology. Conversely, malignant lesions showed lower values, consistent with their irregular and infiltrative structure. These results underscore the potential of shape-based analysis as an interpretable and physiologically meaningful approach to improving breast lesion classification in ultrasound imaging.
Keywords: Breast Cancer Detection, Ultrasound Imaging, Shape Factor Analysis, PLS-DA, Benign and Malignant Lesions
