We present an extension of the self-supervised outlier detection (SSD) framework to the three-dimensional case. We first apply contrastive learning on a network using a general dataset of two-dimensional slices randomly sampled from all the available training data. This network serves as a latent embedding encoder of the input images. We model the in-distribution latent density as a multivariate Gaussian, fitted to the embeddings of the training slices. At test time, each test sample is scored by summing the Mahalanobis distances from all its slices to the means of the learned Gaussians. While mainly meant as a sample-level method, this approach additionally enables coarse localization, scoring each voxel by the minimum Mahalanobis distance among the slices that contain it. On the sample-level task of the 2021 MICCAI Medical Out-of-Distribution Analysis Challenge, our method ranked second on the challenging abdominal dataset, and fourth overall. Moreover, we show that with pretrained features and the right choice of architecture, a further boost in performance can be gained.
