A Quantitative Evaluation Framework for Explainable AI in Semantic Segmentation

Ensuring transparency and trust in artificial intelligence (AI) models is essential as they are increasingly deployed in safety-critical and high-stakes domains. Explainable AI (XAI) has emerged as a promising approach to address this challenge; however, the rigorous evaluation of XAI methods remains vital for balancing the trade-offs between model complexity, predictive performance, and interpretability. While substantial progress has been made in evaluating XAI for classification tasks, strategies tailored to semantic segmentation remain limited. Moreover, objectively assessing XAI approaches is difficult, since qualitative visual explanations provide only preliminary insights. Such qualitative methods are inherently subjective and cannot ensure the accuracy or stability of explanations. To address these limitations, this work introduces a comprehensive quantitative evaluation framework for assessing XAI in semantic segmentation, accounting for both spatial and contextual task complexities. The framework systematically integrates pixel-level evaluation strategies with carefully designed metrics to yield fine-grained interpretability insights. Simulation results using recently adapted class activation mapping (CAM)-based XAI schemes demonstrate the efficiency, robustness, and reliability of the proposed methodology. These findings advance the development of transparent, trustworthy, and accountable semantic segmentation models.