Rare disease classification via difficulty-aware meta learning

Abstract

<p>Deep convolutional neural networks (ConvNets) have achieved state-of-the-art performance in various medical image analysis tasks. The success is partially attributed to a large amount of labeled data. However, some medical images, such as a rare disease case actinic keratosis, are rare and are difficult to obtain a large amount of labeled data in hospitals. Hence, how to train a network to classify medical images in a highly low-data regime, <em>i.e.</em>, <em>three or five samples per class</em>, is quite important but currently catches little attention. This book chapter presents a difficulty-aware meta learning method to address rare disease classification with an application on dermoscopy images. The key idea is to train a model on a variety of learning tasks, such that it can solve new tasks, <em>i.e.</em>, rare disease classification, using only a few labeled samples. The method is motivated by recent progress in meta learning, but differently, we introduce the difficulty-aware meta learning optimization loss (DAML). Our novel difficulty-aware metaoptimization loss is inspired by the observations that the contribution of different task samples is various. By dynamically monitoring the scaling factor, our method can downweight the well-learned tasks and rapidly focus on the hard tasks. We evaluate our method on the ISIC 2018 skin lesion classification dataset. The model can quickly adapt to classify unseen classes with only five samples per class by a high AUC of 83.3%. We also show examples that our method can be used in real clinical practice for rare disease classification.<span> </span></p>