Deep Metric Learning Based Histopathological Image Classification

Abstract

With the rapid development of deep learning, histopathological image classification models have made remarkable progress. Recent deep learning-based methods have been applied to raw histopathological images to construct end-to-end models, which avoid hand-craft feature engineering. To construct a model that can capture the intrinsic pattern of the histopathological image dataset, we design a model based on deep metric learning which embeds data points into a Euclidean space. The proposed model trains a deep neural network, which embeds an input image into a Euclidean space where dissimilar images are located far away to each other and vice versa. We adopt a BN-Inception network pretrained on ImageNet as the embedding model. Then it is retrained on target datasets with some triplet loss function. A weighted distance-based triplet sampling strategy is designed to generate hard triplets for the training procedure. Evaluations on benchmark datasets indicate that our deep metric learning-based method outperforms recent successful deep learning models.