Keywords: Uncertainty estimation, Mixup, Data augmentation, Skin lesion diagnosis
Abstract: Uncertainty is considered to be an important measure that provides valuable information on the learning behavior of deep neural networks. In this paper, we propose an uncertainty estimation method using test-time mixup augmentation (TTMA). The TTMA uncertainty is obtained by replacing affine augmentation with the mixup in the existing test-time augmentation (TTA) method. In addition to the data uncertainty, we propose TTMA-based class-specific uncertainty, which can provide information on between-class confusion. In experiments on the skin lesion diagnosis dataset, we confirmed that the proposed TTMA not only provides better epistemic uncertainty than TTA but also provides information on between-class confusion through class-specific uncertainty.
Paper Type: methodological development
Primary Subject Area: Uncertainty Estimation
Secondary Subject Area: Interpretability and Explainable AI
Paper Status: original work, not submitted yet
Source Code Url: TBD
Data Set Url: https://challenge2018.isic-archive.com/
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.