Deep feature learning for histopathological image classification of canine mammary tumors and human breast cancer

• Published in: Information sciences 2020-01, Vol.508, p.405-421
• Main Authors: Kumar, Abhinav, Singh, Sanjay Kumar, Saxena, Sonal, Lakshmanan, K., Sangaiah, Arun Kumar, Chauhan, Himanshu, Shrivastava, Sameer, Singh, Raj Kumar
• Format: Article
• Language: English
• Subjects: Breast cancer; Canine mammary tumor (CMT); Deep learning; Histopathological classification
• ISSN: 0020-0255; 1872-6291
• DOI: 10.1016/j.ins.2019.08.072

•A CMTHis dataset comprising Canine Mammary Tumor (CMT) histopathological images was introduced.•A framework based on VGGNet-16 for deep feature extraction, along with different classifiers, was proposed and evaluated on CMTHis and BreakHis datasets.•The framework achieved 97% and 93% test accuracy for binary classification of human breast cancer and CMT, respectively. Canine mammary tumors (CMTs) have high incidences and mortality rates in dogs. They are also considered excellent models for human breast cancer studies. Diagnoses of both, human breast cancer and CMTs, are done by histopathological analysis of haematoxylin and eosin (H&E) stained tissue sections by skilled pathologists: a process that is very tedious and time-consuming. The existence of heterogeneous and diverse types of CMTs and the paucity of skilled veterinary pathologists justify the need for automated diagnosis. Deep learning-based approaches have recently gained popularity for analyzing histopathological images of human breast cancer. However, so far, due to the lack of any publicly available CMT database, no studies have focused on the automated classification of CMTs. To the best of our knowledge, we have introduced for the first time a dataset of CMT histopathological images (CMTHis). Further, we have proposed a framework based on VGGNet-16, and evaluated the performance of the fused framework along with different classifiers on the CMT dataset (CMTHis) and human breast cancer dataset (BreakHis). We also explored the effect of data augmentation, stain normalization, and magnification on the performance of the proposed framework. The proposed framework, with support vector machines, resulted in mean accuracies of 97% and 93% for binary classification of human breast cancer and CMT respectively, which validates the efficacy of the proposed system.