Classification of Breast Cancer Images Using Ensembles of Transfer Learning

Abstract

It is a challenging task to estimate the cancerous cells and tissues via computer-aided diagnosis systems on high-resolution histopathological images. In this study, it is suggested to use transfer learning and ensemble learning methods together in order to reduce the difficulty of this task and better diagnose cancer patients. In the studies, histopathological images with 40× and 100× magnification factors are analyzed. In order to prove the success of the study with experimental studies, firstly, the results provided by pre-modeled deep learning architectures trained by histopathological image dataset, then the results acquired by different transfer learning approaches and the results obtained with the ensembles of deeply learned features using transfer learning methods are presented comparatively. Three different approaches are applied for transfer learning by fine-tuning the pre-trained convolution neural networks. In the experimental section, results of single classifiers (i.e., support vector machines, logistic regression, k-nearest neighbor and bagging) are presented by employing features of CNN models obtained by defined transfer learning approaches. Then, decisions of each classifier model are combined separately by weighted decision fusion (WDF) and stacking decision fusion (SDF) ensemble learning methods that have proven to improve the classification performance of the proposed classification system.