Clinically Validated XAI for Calcified Plaque Segmentation in Coronary CT Angiography

Background: Accurate segmentation of calcified plaques in coronary computed tomography angiography (CCTA) images is critical for reliable assessment of coronary atherosclerotic burden and for supporting interpretation of luminal stenosis, yet it remains a challenging task due to annotation inconsistencies, blooming artifacts and low contrast at lesion boundaries. These limitations may affect both automated model performance and the clinical trustworthiness of AI systems. This study explores the impact of annotation refinement on segmentation performance and model explainability, as well as the influence of representation learning on explanation quality. Methods: We trained a deep convolutional neural network to segment calcified plaques in coronary arteries using a dataset of expert-labeled CT slices. Initial training on radiologist-provided annotations yielded suboptimal results. To address this, annotations were manually revised and validated by radiologists. In addition to a standard ImageNet-pretrained model, we evaluated a self-supervised representation learning approach using DINOv2. Grad-CAM was used to generate visual explanations for model predictions before and after annotation refinement. Results: Models trained with refined annotations achieved notably improved segmentation accuracy, with clearer delineation of calcified regions. Grad-CAM localization analysis demonstrated improved concentration of model attention within plaque and vessel regions. Furthermore, models incorporating DINOv2 representations produced more spatially coherent attention maps, with improved anatomical localization consistent with coronary vessel regions and reduced off-target activations, as qualitatively confirmed by expert radiologists. Conclusions: Our findings emphasize the importance of high-quality, validated annotations in developing accurate and interpretable AI models for medical imaging. In addition, the results suggest that representation learning influences the reliability and clinical relevance of explainability outputs. The combination of manual annotation refinement, expert validation, and improved feature representations provides a practical workflow for human-in-the-loop AI development in cardiovascular imaging. This study demonstrates that annotation quality is a critical and often underestimated determinant of XAI reliability, and suggests that explainability methods can serve as feedback tools for iterative, clinician-guided dataset curation in cardiovascular imaging.