Transfer learning for cell nuclei classification in histopathology images

Abstract

In histopathological image assessment, there is a high demand to obtain fast and precise quantification automatically. Such automation could be beneficial to find clinical assessment clues to produce correct diagnoses, to reduce observer variability, and to increase objectivity. Due to its success in other areas, deep learning could be the key method to obtain clinical acceptance. However, the major bottleneck is how to train a deep CNN model with a limited amount of training data. There is one important question of critical importance: Could it be possible to use transfer learning and fine-tuning in biomedical image analysis to reduce the effort of manual data labeling and still obtain a full deep representation for the target task? In this study, we address this question quantitatively by comparing the performances of transfer learning and learning from scratch for cell nuclei classification. We evaluate four different CNN architectures trained on natural images and facial images.