Explainable Artificial Intelligence for Image Quality Assessment

Abstract

The "black-box" nature of many Artificial Intelligence (AI) systems used in Image Quality Assessment (IQA) limits their trustworthiness, especially in critical applications like medical imaging. This thesis introduces a novel framework for eXplainable Artificial Intelligence (XAI) tailored to IQA, demonstrated by detecting and explaining foreign objects in medical X-rays. Our approach integrates a DeepLabV3+ (ResNet50 backbone) semantic seg- mentation model with gradient-based XAI methods (Grad-CAM, NormGrad) to provide visual explanations. A key contribution is an advanced Visualization and Scoring Engine that processes model outputs and saliency maps to derive nuanced image quality scores based on per-object characteristics (size, location, model confidence) and generates Large Language Model (LLM)-based textual summaries of the assessment. Evaluated on the Object-CXR dataset via 5-Fold Cross-Validation, the frame- work demonstrated robust performance. Significantly, XAI applied to the segmenta- tion architecture yielded vastly superior explanation localization (e.g., Grad-CAM Pointing Game Accuracy ≈ 0.512) compared to classification-based baselines (PGA ≈ 0.1 − 0.2). Grad-CAM was found to be more efficient and faithful for this IQA task. The LLM summaries effectively enhanced interpretability. This research presents an end-to-end XAI-IQA system, offering a pathway to more transparent and reliable automated image quality assessment. The developed frame- work and experimental code are publicly available, fostering reproducibility and further research, at https://github.com/MoncefDj/Explainable-Artificial- Intelligence-for-Image-Quality-Assessment.