Publication Date
2026
Document Type
Dissertation/Thesis
First Advisor
Alhoori, Hamed
Degree Name
Ph.D. (Doctor of Philosophy)
Legacy Department
Department of Computer Science
Abstract
Anomaly detection is critical across industrial and healthcare domains, yet the definition of what constitutes an anomaly is both subjective and application-specific. In domains such as non-destructive inspection (NDI) and medical diagnostics, the effectiveness of AI models depends not only on their precision but also on their interpretability, scalability, and ability to operate under data scarcity. This dissertation developed three AI frameworks that address these challenges in distinct application domains.
First, an explainable AI framework for COVID-19 diagnosis using MALDI-ToF mass spectrometry data was developed. A Random Forest classifier trained on area-under-the-curve ratio features achieved 94.12% test accuracy, and a four-stage X-AI pipeline integrating local and global explanations distilled complex feature attributions into clinician-friendly representations.
Second, a sparse Mixture-of-Experts (MoE) architecture with a novel multi-level loss formulation was introduced for defect classification in multi-domain ultrasonic testing data from carbon fiber-reinforced polymer aerospace components. Top-1 sparse routing achieved 91.3% accuracy and 0.869 F1-score, converging at epoch 33 versus epoch 321 for the baseline (approximately 10× speedup).
Third, PRISM (Patch Reconstruction with Integrated Scoring Method), a hybrid generative discriminative framework for visual anomaly detection, was presented. Operating entirely in the feature space of a pretrained backbone with a compact self-attention denoiser (∼2.7M parameters), PRISM achieved 98.4% average image AUROC and 98.3% average pixel AUROC across all 15 MVTec AD categories without any external data.
Together, these contributions demonstrate that annotation-efficient, interpretable, and domain-aware AI design is achievable across diverse modalities, advancing both applied and foundational aspects of artificial intelligence.
Recommended Citation
Seethi, Venkata Devesh Reddy, "Advancing Anomaly Inspection with AI in the Wild: Trustworthy, Scalable, and Generative Frameworks" (2026). Graduate Research Theses & Dissertations. 8244.
https://huskiecommons.lib.niu.edu/allgraduate-thesesdissertations/8244
Extent
151 pages
Language
en
Publisher
Northern Illinois University
Rights Statement
In Copyright
Rights Statement 2
NIU theses are protected by copyright. They may be viewed from Huskie Commons for any purpose, but reproduction or distribution in any format is prohibited without the written permission of the authors.
Media Type
Text
