Prototype transfer generative adversarial network for unsupervised breast cancer histology image classification. (July 2021)

• Record Type: Journal Article
• Authors: Wang, Dan; Chen, Zhen; Zhao, Hongwei
• Is Part Of: Biomedical signal processing and control. Volume 68(2021)
• Publication Date: 2021-07
• Subjects: 00-01 -- 99-00; Image classification -- Breast cancer -- Prototype transfer -- Generative adversarial network; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102713
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Propose a prototype transfer generative adversarial network for unsupervised breast cancer histology image classification task. Develops a target-biased generative adversarial network to transform the source images into target style. Introduces a prototype transfer module to conduct feature-level transfer learning. Abstract: Breast cancer (BC) has become a common tumor that threatens women's health. The decision on the treatment for breast cancer depends on multi-classification. Therefore, for preventive diagnosis, the development of automatic malignant BC detection system suitable for patient imaging can reduce the burden on pathologists and help avoid misdiagnosis. At present, most of the research methods are supervised learning methods that require lots of labeled data, and annotating histology images is more difficult and expensive due to the complicated disease representation in breast cancer. In this paper, we propose an unsupervised learning method, named prototype transfer generative adversarial network (PTGAN), which embeds generative adversarial networks and prototypical networks for classifying a large number of data sets by training a transfer learning model from a small number of labeled source data sets from similar domain. Without requiring lots of labeled target images, this method also reduces the style difference between the source and target domains by generating an adversarial network, thereby it can effectively reduce the pixel-level distribution gap for breast histology images captured from different devices with individual style. Then, it embeds the feature vectors learned by a prototype network into the metric space, which can distil discriminative knowledge from the prototype into target domain. We then use a special "distance" in the metric space to train a classifier to predict the large amounts of target data. The experimental results on the BreakHis dataset show that the accuracy of the proposed PTGAN for classifying benign and malignant tissues has reached nearly 90%. This proves the advantage of our method in providing an effective tool for breast cancer multi-classification in clinical settings, economizing the complicated annotating cost.