A deep learning approach to automate human-visual inspections in the medical-devices manufacturing industry

This thesis investigates how deep learning can be used to automate visual inspection in medical device manufacturing, a domain where extremely high accuracy is required and manual inspection remains prevalent despite its susceptibility to fatigue, variability, and error. Traditional rule-based machine vision performs well for dimensional checks but is poorly suited to attribute inspections that depend on subtle, context-dependent visual cues. To address this gap, the research focuses on methods that can be trained predominantly on defect-free samples, reflecting the rarity and cost of obtaining representative defective parts in highly regulated environments. The work makes four main contributions. First, it introduces and publicly releases a dataset of 1,200 annotated images of sterile barrier packaging with realistic seal defects, enabling reproducible benchmarking. Second, it reports a large-scale evaluation of 27 convolutional neural network architectures across multiple small and imbalanced dataset regimes, and proposes the median distance to the boundary as a robustness metric for detecting overfitting when test accuracies approach 100%. Third, it develops two attention-guided autoencoder architectures for deep anomaly detection: a structural-similarity variant tailored for lightweight, real-time inspection on production lines, and a feature-distance variant designed for lifecycle monitoring in supervisory systems; both are validated on an external dataset and shown to be competitive with state-of-the-art anomaly-detection approaches. Finally, the thesis provides a technical and regulatory analysis of the EU Artificial Intelligence Act and related medical-device frameworks, highlighting challenges in data governance, explainability, validation, and post-deployment monitoring. Overall, the findings demonstrate that deep learning models trained mainly on defect-free data can deliver scalable, high-performing and more easily qualifiable automated inspection systems, while also identifying open issues that motivate future research at the intersection of advanced manufacturing and AI regulation.