Soutenance publique de thèse de doctorat en informatique - Arnaud BOUGAHAM

Défense publique de thèse de doctorat

This thesis proposes a robust, interpretable, and transferable deep-learning framework for anomaly detection in safety-critical domains such as industrial quality control and medical diagnostics. These two fields, though distinct, share major challenges: class imbalance with limited abnormal samples, and the need for trustworthy and real-time decisions under strict reliability constraints. The main objective is, thus, to build methods with minimal supervision (training with normal data only) while allowing a human-aligned interpretability.

The approach combines artificial intelligence unsupervised generative modeling with supervised classification, focusing on patching techniques, local representation, and interpretable scoring. Four key contributions structure this work:
(i) Generative Adversarial Network Anomaly Detection through Intermediate
Patches (GanoDIP), a Generative Adversarial Network (GAN) architecture for high-resolution, industrial anomaly localization at the patch level.
(ii) Vector Quantized Generative Adversarial Network Anomaly Detection
through Intermediate Patches (VQGanoDIP), an extension with vector-quantized latent representations and composite scoring for improved reconstruction and fidelity.
(iii) Cycle Generative Adversarial Network-Anomaly Detection (CGAN-AD), a conditional image translation model that integrates both normal and abnormal data for enhanced domain transfer in industrial and medical settings.
(iv) Trustworthy approximated partial AUC (tapAUC), a loss function that enforces the Zero False Negative constraint, for high recall in critical scenarios.

These models are deployed in real-world use cases. The framework is integrated into an active production line to detect unexpected components (such as screws) in printed circuit boards, delivering interpretable decisions with minimal false alarms. In the medical domain, the approach is adapted to well perform in Positron Emission Tomography (PET) based coma receptivity analysis, and some techniques are incorporated for an ovarian cancer segmentation application. These results demonstrate not only technical efficiency but also organizational viability through human-in-the-loop deployment and real-time scalability.

Together, these contributions establish a modular, constraint-aware, and explainable anomaly detection paradigm, advancing the field towards trustworthy and human-centered artificial intelligence in high-stakes environments.

Contact : Sara Medugno - 081 72 49 92 - sara.medugno@unamur.be
Télecharger : vCal