Integrating multiscale features for robust breast cancer histopathology image classification

Abstract

Breast cancer is a prominent contributor to cancer-related fatalities among women globally, and early detection via histopathological examination, despite being regarded as the benchmark method, is frequently time-consuming and prone to subjectivity. To overcome these challenges, we introduce an automated deep learning framework that categorizes Hematoxylin and Eosin (H&E)-stained breast tissue samples into four groups: normal tissue, benign lesions, localized malignant tumors, and invasive carcinoma. Our system utilizes intermediate feature embeddings from the Xception (Extreme Inception) architecture to derive discriminative features, attaining 98% precision, a 0.969 kappa metric, and exceptional AUC-ROC 0.998 and AUC-PR 0.995 values on the raw (non-normalized) dataset, with heightened detection rates for localized malignant tumors 96% and invasive carcinoma, 99%. We also assessed four stain standardization methods (Reinhard, Ruifrok, Macenko, and Vahadane) and found that Macenko normalization yielded the strongest results: 97.79% accuracy, kappa = 0.965, AUC-ROC = 0.997, AUC-PR = 0.991. However, unprocessed images still outperformed all normalized versions. A comparative study revealed that our method exceeds baseline AlexNet and state-of-the-art deep learning architectures (VGG16, VGG19, Inception-v3, and conventional Xception), highlighting its capability to improve diagnostic reliability while confirming that stain normalization, though viable, does not surpass the effectiveness of the unaltered dataset.