Spot the Relevance: Evaluating Pixel Attribution Methods for Clinical Significance in Radiological Imaging

Pixel Attribution Methods (PAMs) are essential techniques in Explainable Artificial Intelligence (XAI) for improving the interpretability of black-box models in diagnostic imaging, particularly through saliency maps that highlight regions of interest in medical images. Despite their potential, existing PAMs often fail to align with clinical reasoning and radiological practices, limiting their applicability in real-world settings. This study addresses this gap by evaluating the semantic significance, defined as relevance and pertinence, of saliency maps generated by different PAMs (Grad-CAM, XGrad-CAM, Grad-CAM++, and Smooth Grad-CAM++) at layers 3 and 4 of a ResNeXt-50 neural network. Using a dataset of 12 thoraco-lumbar X-rays, we quantified the alignment between AI-generated saliency maps and human-defined maps created by six specialists, which were then aggregated using various criteria. Among the evaluated methods, Smooth Grad-CAM++ exhibited the lowest performance, while Grad-CAM, XGrad-CAM, and Grad-CAM++ yielded comparable results. Furthermore, saliency maps from layer 4 demonstrated significantly higher relevance than those from layer 3, although the difference in pertinence was less pronounced. However, saliency maps consistently underperform compared to expert ground truth annotations. These findings underscore the limited clinical significance of current saliency map approaches, highlighting the need for more human-centered AI models that better integrate into real-world diagnostic decision-making processes.