Fine-Grained Out-of-Distribution Detection of Medical Images Using Combination of Feature Uncertainty And Mahalanobis Distance

Automatic medical image classification methods are often faced with out-of-distribution (OOD) samples belonging to disease classes that are unseen during training in clinical practice, and incorrectly predict these samples into one of the categories in the training dataset, i.e, in-distribution (ID) classes, leading to unreliable outputs. Therefore, it is essential to detect OOD samples for reliable prediction of models. To address this issue, we propose a simple yet effective approach that combines the uncertainty of multi-scale features of a test sample and the Mahalanobis distance between the test sample and distributions of training classes in the feature space through logistic regression. We evaluated our approach on a dataset containing five fine-grained lung conditions and designed three ID-OOD splits. The experimental results averaged over the three splits showed that our method outperformed existing methods in detecting OOD samples in medical images.