Leveraging Neural Networks and Advanced LBP Features by Performance Driven Breast Cancer Detection

Abstract

Breast cancer detection from mammography images is a critical task in medical diagnostics, requiring precise and robust feature extraction techniques. In this study, we present a quantitative analysis of various advanced Local Binary Pattern (LBP) techniques for enhancing feature extraction in breast cancer detection from mammography images, leveraging deep learning models. We systematically evaluate uniform, rotation-invariant, extended, completed, multi-scale, elliptical, and volume LBP features, alongside other texture descriptors such as Histogram of Oriented Gradients (HOG), Gabor filters, and Gray Level Co-occurrence Matrix (GLCM). We use the Digital Database for Screening Mammography (DDSM) to assess the impact of these feature extraction methods on classification accuracy with deep neural networks. Our analysis spans multiple neural network architectures, including AlexNet, VGG-19, EfficientNetB0, MobileNet, and ResNet50. We employ stratified k-fold cross-validation to ensure statistical robustness and data partitioning consistency. The empirical framework employs rigorous training and generalization callbacks, such as early stopping and learning rate optimizers. We conduct statistical hypothesis testing, specifically ANOVA, to determine the significance of performance variations among different features. Results indicate that complete LBP features significantly enhance mean classification accuracy by 7.13% across all models, with EfficientNetB0 achieving a maximum improvement of 9.41%. Our statistical analysis confirms that the observed performance enhancements are statistically significant, validating the superiority of complete LBP over other image features. These findings underscore the mathematical efficacy of advanced LBP techniques in augmenting deep learning models' performance for breast cancer detection from mammography images.