


Background:
Medical Out-of-Distribution Analysis Challenge - MICCAI 2023
Medical imaging methods need to address the issue of OOD data, however many methods developed for their identification perform sub-optimally with real-world data. This thesis examines how disentangling latent spaces can lead to improved OOD detection on benchmarking challenges including MICCAI 2023 and existing datasets, as well as new datasets. This thesis will examine their advantages and limitations in a systematic manner. OOD data will include: (1) data with artificially introduced noise (from Gaussian to more realistic artifacts), (2) MICCAI 2023 challenge data, (3) benchmarking cases from the literature (e.g. [1]). It will also consider alternative, low-cost OOD metrics [3], to assess their reliability and propose robust alternatives.
Resources: