Dual-Mode Deep Anomaly Detection for Medical Manufacturing: Structural Similarity and Feature Distance

Automated visual inspection in medical device manufacturing faces unique challenges, including small and imbalanced datasets, high-resolution imagery, and strict regulatory requirements. To address these, we propose two attention-guided autoencoder architectures for deep anomaly detection. The first employs a structural similarity-based scoring approach that enables lightweight, real-time defect detection with unsupervised thresholding and can be further enhanced through limited supervised tuning. The second applies a feature distance-based strategy using Mahalanobis scoring on reduced latent features, designed to monitor distributional shifts and support supervisory oversight. Evaluations on a representative sterile packaging dataset confirm that both approaches outperform baselines under hardware-constrained, regulated conditions. Cross-domain testing on the MVTec-Zipper benchmark further demonstrates that the structural similarity-based method generalises effectively and achieves performance comparable to state-of-the-art methods, while the feature distance-based method is less transferable but provides complementary monitoring capabilities. These results highlight a dual-pathway inspection strategy: structural similarity for robust inline detection and feature distance for supervisory monitoring. By combining operational performance with interpretability and lifecycle monitoring, the proposed methods also align with emerging regulatory expectations for high-risk AI systems.