Stepwise Transfer of Domain Knowledge for Computer-Aided Diagnosis in Pathology Using Deep Neural Networks

Abstract

Deep learning using deep convolutional neural networks (DCNNs) has demonstrated unprecedented power in image classification. Subsequently, computer-aided diagnosis (CAD) for pathology imaging has been largely facilitated by DCNN approaches. However, because DCNNs require massive amounts of labeled data, the lack of availability of such pathology image data as well as the high cost of labeling new data is currently a major problem. To avoid expensive data labeling efforts, transfer learning is a concept intended to overcome training data shortages by passing knowledge from the source domain to the target domain. However, weak relevance between the source and target domains generally leads to less effective transfers. Therefore, following the natural step-by-step process by which humans learn and make inferences, a stepwise fine-tuning scheme is proposed by introducing intermediate domains to bridge the source and target domains. The DCNNs are expected to acquire general object classification knowledge from a source domain dataset such as ImageNet and pathology-related knowledge from intermediate domain data, which serve as fundamental and specific knowledge, respectively, for the final benign/malignant classification task. To realize this, we introduce several ways to provide pathology-related knowledge by generating an intermediate dataset classified into various corresponding pathology features. In experiments, the proposed scheme has demonstrated good performance on several well-known deep neural networks.