A Feature-Based Out-of-Distribution Detection Approach in Skin Lesion Classification

• Thiago Carvalho PUC-RJ https://orcid.org/0000-0001-8689-1438
• Marley Vellasco PUC-RJ https://orcid.org/0000-0002-9790-1328
• José Franco Amaral UERJ https://orcid.org/0000-0003-4951-8532
• Karla Figueiredo UERJ https://orcid.org/0000-0001-8420-3937

Resumo

When dealing with Deep Learning applications in open-set problems, accurately classifying known classes seen in the training phase is not the only aspect to be taken into account. In such a context, detecting Out-of-Distribution (OOD) samples plays an important role as an auxiliary task, generally solved by OOD detection methods. For medical applications, detecting unknown samples may in classification problems can be beneficial for many aspects, such as a better understanding of the diagnosis and probably a more adequate treatment. In this article, we evaluate a feature space-based approach, named as OpenPCS-Class, for OOD detection in medical applications, more specifically skin lesion classification. We compare the OpenPCS-Class against important OOD detection methods, evaluating different model architectures and OOD datasets. The OpenPCS-Class outperformed other methods at 48.4% and 5.3% in terms of FPR95 and AUROC, respectively.