Abstract
Medical anomaly detection is challenged by subtle pathologies, limited abnormal training samples, and modality heterogeneity. Reconstruction-based autoencoders are widely used in anomaly detection, but may fail when anomalous inputs are reconstructed too accurately. In this paper, we propose a dual-criterion anomaly detection approach that combines image-space reconstruction evidence with latent-space deviation, using Adversarial Autoencoders. The approach was validated on seven heterogeneous benchmark datasets covering 2D and 3D medical images. Comparative evaluation against alternative autoencoder variants showed that AAE achieved the strongest average performance, with an aggregate score of 0.723. These results support the use of adversarially regularized autoencoders as a suitable foundation for dual-criterion anomaly detection in heterogeneous medical images.
Paper Type
Poster
DOI
10.62036/ISD.2026.60
Dual-Criterion Anomaly Detection in Heterogeneous Medical Images Using Adversarial Autoencoders
Medical anomaly detection is challenged by subtle pathologies, limited abnormal training samples, and modality heterogeneity. Reconstruction-based autoencoders are widely used in anomaly detection, but may fail when anomalous inputs are reconstructed too accurately. In this paper, we propose a dual-criterion anomaly detection approach that combines image-space reconstruction evidence with latent-space deviation, using Adversarial Autoencoders. The approach was validated on seven heterogeneous benchmark datasets covering 2D and 3D medical images. Comparative evaluation against alternative autoencoder variants showed that AAE achieved the strongest average performance, with an aggregate score of 0.723. These results support the use of adversarially regularized autoencoders as a suitable foundation for dual-criterion anomaly detection in heterogeneous medical images.
Recommended Citation
Wosiak, A., Stępniak, M. & Żykwińska, K.(2026). Dual-Criterion Anomaly Detection in Heterogeneous Medical Images Using Adversarial Autoencoders. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.60