Magnification Independent Multi-Classification of Breast Cancer in Histopathology Images Using Deep Learning
Jason Nishio, Nelson Nishio, Alexander Nishio · 2024
Breast cancer is one of the most common and leading causes of female mortality in the world. The disease can be treated quickly and successfully if the tumor type is diagnosed at an early stage. Histopathology provides superior breast cancer analysis. However, due to tumor cell heterogeneity, manually identifying and classifying breast tumors in histopathological images is highly laborintensive for pathologists and can sometimes be inaccurate. Hence, there is a need to develop an automated detection framework using deep learning. This study leverages the pretrained ResNet50 and VGG19 models for detecting breast cancer in histopathological images from the BreakHis dataset. Unlike previous research that often separated images by magnification for classification, this study bridges this gap by constructing models for magnification-independent multi-tumor-type classification. The key success of this study is the proposed Z-score color normalization in RGB channels during image preprocessing, which reduces variations in color, brightness, and contrast. The pretrained ResNet-50 and VGG19 models achieved 99.03% and 96.59% accuracies in binary classification, and 94.73% and 92.50% accuracies in multiclassification, respectively. These results surpass the performance of existing methods in magnification-independent multiclassification for breast cancer detection.