📝 SummarXiv

Abstract

Automating visual inspection in medical device manufacturing remains challenging due to small and imbalanced datasets, high-resolution imagery, and stringent regulatory requirements. This work proposes two attention-guided autoencoder architectures for deep anomaly detection designed to address these constraints. The first employs a structural similarity-based anomaly score (4-MS-SSIM), offering lightweight and accurate real-time defect detection, yielding ACC 0.903 (unsupervised thresholding) and 0.931 (supervised thresholding) on the - Surface Seal Image - Test split with only 10% of defective samples. The second applies a feature-distance approach using Mahalanobis scoring on reduced latent features, providing high sensitivity to distributional shifts for supervisory monitoring, achieving ACC 0.722 with supervised thresholding. Together, these methods deliver complementary capabilities: the first supports reliable inline inspection, while the second enables scalable post-production surveillance and regulatory compliance monitoring. Experimental results demonstrate that both approaches surpass re-implemented baselines and provide a practical pathway for deploying deep anomaly detection in regulated manufacturing environments, aligning accuracy, efficiency, and the regulatory obligations defined for high-risk AI systems under the EU AI Act.