Date of Graduation
Spring 2026
Degree
Master of Science in Computer Science
Department
Computer Science
Committee Chair
Rahul Dubey
Abstract
Deep learning models have become central to image analysis across a wide range of domains. However, their black-box nature raises critical concerns about interpretability, stability, and trust in decision-making. In particular, domains such as artificial intelligence in healthcare highlight the critical need for reliable and interpretable model behavior, given their direct impact on high-stakes decision-making. Prior research has shown that widely used Explainable Artificial Intelligence (XAI) methods can highlight relevant regions that influence model decisions. But they often suffer from stochastic inconsistency, segmentation sensitivity, and limited control over the trade-off between explanation compactness and faithfulness, motivating the need for more robust explanation frameworks. To address these challenges, this thesis proposes two methodologies: (1) aggregation-based approach to improve consistency in a well-known perturbation based XAI technique; Local Interpretable Model Agnostic Explanation (LIME) and (2) an optimization-based explanation technique. These methods were evaluated across diverse datasets, including both natural image benchmarks and medical imaging datasets. We considered both qualitative and quantitative evaluation metrics for benchmarking against other state-of-the-art XAI methods. Experimental results show that the proposed aggregation-based explanation method performed statistically significantly better than the state-of-the-art XAI methods. Additionally, the explanation was cast as a multi objective problem to maximize both fidelity and compactness. Experiments were conducted considering different image segmentation methods to observe changes in the explanations and for further analysis. The findings demonstrate improved stability, stronger faithfulness, and better compactness–fidelity trade-offs compared to baseline methods such as LIME, contributing to a more principled and clinically reliable foundation for explainable AI.
Keywords
explainable artificial intelligence, medical imaging, perturbation-based explanations, multi-objective optimization, superpixel segmentation
Subject Categories
Computer Engineering | Health Information Technology
Copyright
© Abiha Tahsin Chowdhury
Recommended Citation
Chowdhury, Abiha Tahsin, "Novel Stability and Optimization Driven Explainable AI for Classification Tasks" (2026). Graduate Theses/Dissertations. 4180.
https://bearworks.missouristate.edu/theses/4180
