Date of Graduation
Summer 2026
Degree
Master of Science in Computer Science
Department
Computer Science
Committee Chair
Rahul Dubey
Abstract
Machine learning is increasingly used in high-stakes domains such as healthcare, finance, and the judiciary, where fairness, transparency, and accountability are essential. In medical imaging, models are expected not only to classify diseases accurately but also to base their decisions on clinically relevant regions. Explainable Artificial Intelligence (XAI) methods are commonly used to generate heatmaps that highlight image regions influencing a model's prediction. However, these explanations are typically generated after training and do not influence how the model learns, allowing models to achieve high accuracy while relying on spurious or non-causal features. This thesis proposes explanation-aware deep learning methodologies that incorporate explanation quality directly into the training process. Models are penalized when their explanation heatmaps focus on regions outside expert-annotated areas, encouraging attention to clinically meaningful features. Experiments were conducted using Convolutional Neural Networks (CNNs) and Transformer architectures on the annotated VinDr Chest X-ray dataset. Results show that the proposed approach improves explanation quality without sacrificing predictive performance and produces models that focus on relevant regions when making predictions. Furthermore, the method generalizes across different model architectures. For binary classification tasks, disease localization improved by approximately 3 to 7 times, while improvements of 2 to 6 times were observed in multi-label classification compared to standard training. Additional experiments using multiple XAI techniques demonstrate that explanation-aware training improves the alignment between model predictions and clinically relevant evidence. These findings establish explanation quality as an explicit optimization objective and provide a practical framework for developing more interpretable and trustworthy biomedical AI systems.
Keywords
Explainable Artificial Intelligence, Explanation-Aware Learning, Biomedical Imaging, Grad-CAM, Convolutional Neural Networks, Vision Transformers
Subject Categories
Artificial Intelligence and Robotics | Computer Sciences
Copyright
© Zubair Faruqui
Recommended Citation
Faruqui, Zubair, "Learning With Explanations: Explanation-Aware Training for Interpretable Medical Imaging Models" (2026). Graduate Theses/Dissertations. 4203.
https://bearworks.missouristate.edu/theses/4203
