Deep Learning Algorithms for Breast Cancer Histopathology Classification in H&E-Stained Images

Breast cancer is a leading cause of cancer deaths worldwide. Diagnosing breast histological images is challenging due to their complexity, compounded by the limited and unevenly distributed availability of pathology experts. This results in a burdensome and time-consuming diagnostic process, with inconsistencies due to subjectivity and varying expertise levels. Therefore, research on computer-aided diagnosis (CAD) is essential. This study aims to assist in distinguishing H&E-stained smear images from various categories of breast cancer tissues by exploring feature extraction, pre-processing, classifier, and input image strategies. The goal is to alleviate the burden on pathologists and standardize interobserver diagnoses. The method developed aims to classify breast cancer into four states and eight morphological subtypes, using feature extractors such as DenseNet201, InceptionV3, and VGG16. Pre-processing methods include data expansion, on-the-fly augmentation, stain normalization, and CLAHE. Input images and classifiers include single input images with different magnifications and multiple patches with majority voting. Analytical methods used were accuracy and loss curves, and evaluation measures included accuracy, precision, and F1-score. The optimal model for 4-class classification used single input images with multiple magnifications, achieving an accuracy, of 95.35%. For breast cancer state classification, the best model used multiple patches with majority voting, achieving an accuracy and F1-score of 85.3% and 85.3%, which improved to 91.25% with majority voting. Despite limitations, this method shows potential for improving tissue analysis and aiding classification. In the future, it may be adapted to help diagnose breast cancer in Indonesia, enhancing diagnostic accuracy and efficiency.