Classification of Breast Cancer Histology Images Using Transfer Learning

Abstract

Breast cancer is one of the leading causes of mortality in women. Early detection and treatment are imperative for improving survival rates, which have steadily increased in recent years as a result of more sophisticated computer-aided-diagnosis (CAD) systems. CAD systems are essential to reduce subjectivity and supplement the analyses conducted by specialists. We propose a transfer learning based approach, for the task of breast histology image classification into four tissue sub-types, namely, normal, benign, in situ carcinoma and invasive carcinoma. The histology images, provided as part of the BACH 2018 grand challenge, were first normalized to correct for color variations induced during slide preparation. Subsequently, image patches were extracted and used to fine-tune Google‘s Inception-V3 and ResNet50 convolutional neural networks (CNNs), both pre-trained on the ImageNet database, enabling them to learn domain-specific features, necessary to classify the histology images. Classification accuracy was evaluated using 3-fold cross validation. The Inception-V3 network achieved an average test accuracy of 97.08% for four classes, marginally outperforming the ResNet50 network, which achieved an average accuracy of 96.66%.