Constructing GLCM-based texture feature expressions using genetic programming for breast tumor classification

Gray-level co-occurrence matrix (GLCM)-based features are commonly utilized in image analysis to extract quantitative textural information. Second-order statistical measures, known as Haralick features, are frequently employed in breast imaging to differentiate between benign and malignant tumors. Due to their general-purpose nature, it is common to calculate many of these features, assuming that some will be relevant to the classification task. However, the broad applicability of these texture features can limit the effectiveness of describing texture patterns unique to breast images. In this paper, GLCM-based texture features are revisited by introducing a genetic programming (GP)-based method to automatically construct more specialized features for breast tumor classification on ultrasound images. This method generates GLCM-based texture feature expressions that lead to low-dimensional feature spaces while maintaining high generalization capability. Compared with 11 reference methods, including feature selection techniques, GP-based feature engineering methods, and convolutional neural networks, the proposed approach demonstrates significantly improved performance across most comparisons. The experimental results demonstrate that generating specific texture features improves classification performance and reduces the dimensionality of the feature space, enabling the proposed method to be potentially applied to other medical imaging classification tasks.