What Do AEs Learn? Challenging Common Assumptions in Unsupervised Anomaly Detection

Abstract

Detecting abnormal findings in medical images is a critical task that enables timely diagnoses, effective screening, and urgent case prioritization. Autoencoders (AEs) have emerged as a popular choice for anomaly detection and have achieved state-of-the-art (SOTA) performance in detecting pathology. However, their effectiveness is often hindered by the assumption that the learned manifold only contains information that is important for describing samples within the training distribution. In this work, we challenge this assumption and investigate what AEs actually learn when they are posed to solve anomaly detection tasks. We have found that standard, variational, and recent adversarial AEs are generally not well-suited for pathology detection tasks where the distributions of normal and abnormal strongly overlap. In this work, we propose MorphAEus, novel deformable AEs to produce pseudo-healthy reconstructions refined by estimated dense deformation fields. Our approach improves the learned representations, leading to more accurate reconstructions, reduced false positives and precise localization of pathology. We extensively validate our method on two public datasets and demonstrate SOTA performance in detecting pneumonia and COVID-19. Code: https://github.com/ci-ber/MorphAEus.