Meta-Learning for Medical Image Classification

Shi Hu, Jakub Tomczak, Max Welling

Abstract: Most of the current state-of-the-art methods to classify medical images is to first train a deep model on ImageNet, then transfer all network weights to a new network except for the last softmax layer, and then fine-tune on the target dataset. When the amount of training data in the target dataset is sufficient, this method is able to surpass the level of a trained doctor on several datasets; however, when it is insufficient, which is common in a lot of real medical applications, this method may lead to mediocre results. To address the small dataset problem, we apply a meta-learning method to train, and then fine-tune on the target dataset. We show our results surpass the state-of-the-art method on a popular medical image dataset.

Author Affiliation: University of Amsterdam

Keywords: deep learning, meta learning, medical image analysis