Handcrafted Feature-based Machine Learning Framework for Robust Breast Cancer Detection in Mammograms

Abstract Background: Early detection of breast cancer plays a crucial role in improving patient survival rates. Methods: In this study, a new pipeline based on novel handcrafted features is developed for automatic classification of mammography images obtained from screening-level examinations, which predominantly represent early-stage and prediagnostic breast findings. The employed features include those extracted from the analysis of images in textural and regional approaches. For texture and region analysis, gray-level co-occurrence matrix (GLCM) and gray-level size zone matrix (GLSZM) are utilized. Moreover, new statistical features are extracted from the input image and its gradient. Three machine learning algorithms, namely XGBoost, LightGBM, and support vector machine, are trained using manual hyperparameter grid search and validated via 10-fold cross-validation. Results: Results demonstrated that the combination of GLCM and GLSZM features, along with appropriate preprocessing, significantly improves the classification accuracy of normal and cancerous images. The proposed method is capable of providing an accuracy of 97.19 for the classification task. Conclusions: This study highlights that integrating textural and statistical features with data balancing techniques can support the development of computer-aided diagnosis systems for breast cancer. The findings suggest that such an approach can provide a cost-effective and interpretable alternative to deep learning models, making it particularly suitable for healthcare centers with limited computational resources. Furthermore, the proposed framework holds the potential for extension to multimodal imaging data, paving the way toward more comprehensive decision-support systems.