Towards Reliable Deep Learning: Feature-Based Out-Of-Distribution Detection for Brain Morphometry

Motivation: Despite recent advancements in image segmentation, supervised deep learning algorithms struggle to generalize to Out-Of-Distribution data. Goal(s): Explore Out-of-Distribution Detection (OODD) in the context of volume-based morphometry for 3D-T1w brain images. Approach: We estimate a training distribution through a patch-based Convolutional-Neural-Network designed for skull-stripping, which extracts essential features from In-Distribution (ID) data. Then, we classify patients (ID vs. OOD) by calculating the distance in feature space between test patches and this established training distribution. Results: Our OODD method correctly classifies 98% of Test-ID subjects and 86% Far-OOD. However, it misclassifies most Near-OOD scans suggesting that the skull-stripping-network alone is insufficient for all use-cases. Impact: We experiment feature-based Out-Of-Distribution (OOD) detection to identify problematic scans for which segmentation results might be unreliable. While Near-OOD remains an area of future improvement, our approach is effective for the majority of use cases and adds negligible computation time.