Typicality Excels Likelihood for Unsupervised Out-of-Distribution Detection in Medical Imaging

Research output: Chapter in Book/Report/Conference proceedingConference contributionAcademicpeer-review

4 Downloads (Pure)

Abstract

Detecting pathological abnormalities in medical images in an unsupervised manner holds potential for advancing modern medical diagnostics. However, supervised methods encounter challenges with exceedingly unbalanced training distributions due to limited clinical incidence rates. Likelihood-based unsupervised Out-of-Distribution (OOD) detection with generative models, especially Normalizing Flows, in which pathological abnormalities are considered OOD, could offer a promising solution. However, research in this direction has shown limited success as prior work has revealed that the likelihood does not accurately reflect the degree of anomaly for OOD samples, where in many instances higher likelihoods are assigned to anomalous samples compared to training samples. In this study, we present the first exploration of typicality (i.e. determining if samples belong to the typical set) for OOD detection in medical imaging, where test samples are juxtaposed against the probability mass rather than the density. The obtained findings demonstrate the superiority of evaluating typicality against likelihood for finding pathological abnormalities. We achieve state-of-the-art performance on the ISIC, COVID-19, and RSNA Pneumonia datasets, while being robust against significant data imbalances.
Original languageEnglish
Title of host publicationUncertainty for Safe Utilization of Machine Learning in Medical Imaging
Subtitle of host publication6th International Workshop, UNSURE 2024, Held in Conjunction with MICCAI 2024, Marrakesh, Morocco, October 10, 2024, Proceedings
EditorsCarole H. Sudre, Raghav Mehta, Cheng Ouyang, Chen Qin, Marianne Rakic, William M. Wells
Place of PublicationCham
PublisherSpringer
Chapter14
Pages149-159
Number of pages11
ISBN (Electronic)978-3-031-73158-7
ISBN (Print)978-3-031-73157-0
DOIs
Publication statusPublished - 3 Oct 2024
Event6th International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, UNSURE 2024 - Marrakesh, Morocco
Duration: 10 Oct 202410 Oct 2024

Publication series

NameLecture Notes in Computer Science (LNCS)
Volume15167
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Workshop

Workshop6th International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, UNSURE 2024
Country/TerritoryMorocco
CityMarrakesh
Period10/10/2410/10/24

Fingerprint

Dive into the research topics of 'Typicality Excels Likelihood for Unsupervised Out-of-Distribution Detection in Medical Imaging'. Together they form a unique fingerprint.

Cite this