Abstract

Microscopic analysis of breast cancer images is the primary task in diagnosing cancer malignancy, which requires high expertise and precision. Recent attempts to automate this highly subjective task have employed deep learning models whose success has depended on large volumes of data while acquiring annotated data in biomedical domains is time-consuming and may not always be feasible. A typical strategy to address this is to apply transfer learning using pre-trained models on a large natural image database (e.g., ImageNet) instead of training a model from scratch. This approach, however, has not been effective in several previous studies due to fundamental differences in patterns, size, and data features between natural and medical images. In this study, we propose and compare several transfer learning approaches that, in the pre-training phase, use both unrelated natural images and related histopathological images to our target data (i.e., cytological images) in order to classify breast cancer cytological biopsy specimens. To our best knowledge, this is the first reported effort to employ a histopathology data source in transfer learning to classify cytological images of breast cancer. Despite intrinsic differences between histopathological and cytological images, we demonstrate that the features learned by the deep networks during the pre-training are compatible with those obtained throughout fine-tuning with the target data set. To thoroughly investigate this assertion, we explore three different strategies for training as well as two different approaches for fine-tuning deep learning models. The proposed method is compared with five state-of-the-art studies previously conducted on the same data set of cytological biopsy images, and we demonstrate that the proposed approach significantly outperforms all of them in terms of classification accuracy. Specifically, the proposed method boasts of improved classification accuracy by 6% to 17% compared to the state-of-the-art studies which were based on traditional machine learning techniques, and also enhanced accuracy by roughly 7% compared to those who utilized deep learning methods, eventually achieving 94.55% test set accuracy and 98.73% validation accuracy. Experimental results show that our approach, despite using a very small number of training images, has achieved performance comparable to that of experienced pathologists and has the potential to be applied in clinical settings.