MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Difficulty-Aware Meta-Learning for Rare Disease Diagnosis

X Li*, L Yu, L Xing, Stanford Univ School of Medicine, Stanford, CA

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Deep convolutional neural networks have shown remarkable breakthroughs in many medical image recognition tasks. The success is partially attributed to a large amount of labeled data. However, how to train a network to classify rare diseases in an extremely low-data regime, i.e., three or five samples per class, catches little attention. In this work, we propose a novel deep learning method to conduct rare disease classification in an extremely low-data regime.

Methods: We present a novel difficulty-aware meta-learning method to address rare disease classifications with an application for the dermoscopy images. The main principle of our method is to train a model on a variety of learning tasks, such that it can solve new tasks, i.e., rare diseases, 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 (DAML). Our novel difficulty-aware meta-optimization is inspired by the observations that the contribution of different task samples is various. By dynamically monitoring the scaling factor, our method can down-weight the well-learned tasks and rapidly focuses on the hard tasks.

Results: We evaluate our method on the ISIC 2018 skin lesion classification dataset. With only five samples in per class, the model can quickly adapt to classify unseen classes by a high AUC of 83.3%. We also validate our method on several rare disease classification tasks using the public Dermo t Image Library1 and achieve a high AUC of 82.67% under the five samples setting, demonstrating the potential of our method for real clinical practice.

Conclusion: We demonstrate that our method can learn from common diseases and then predict classification on the unseen diseases during inference. This technique could be useful for rare disease classification, where the diseases have only few samples.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email