Out-of-Distribution Detection for Skin Lesion Images with Deep Isolation Forest

摘要

In this paper, we study the problem of out-of-distribution (OOD) detection in skin lesion images. Publicly available medical data-sets have a limited number of lesion classes compared to the number of possible diseases in real-life clinical applications. It is thus essential to develop methods that leverage available disease classes in existing datasets to detect previously-unseen types in an unsupervised manner. Toward this goal, we propose an unsupervised and non-parametric OOD detection approach, called DeepIF, which learns the normal distribution of features in a pre-trained CNN using Isolation Forests. We conduct comprehensive experiments on two different datasets and compare our DeepIF against four baseline models. Results demonstrate state-of-the-art performance of our proposed approach on the task of detecting unseen skin lesions.