Clinically Validated XAI for Calcified Plaque Segmentation in Coronary CT Angiography

Area of Science:

• Medical Imaging AI
• Cardiovascular Imaging
• Explainable AI (XAI)

Background:

• Accurate segmentation of coronary calcified plaques in CCTA is vital for assessing atherosclerotic burden and stenosis.
• Challenges include annotation inconsistencies, artifacts, and low contrast, impacting AI model performance and trust.
• This study investigates annotation refinement and representation learning's effect on AI segmentation and explainability.

Purpose of the Study:

• To evaluate the impact of refined annotations on AI segmentation accuracy for coronary calcified plaques.
• To assess how representation learning influences the quality and clinical relevance of AI model explanations.
• To explore a practical workflow for human-in-the-loop AI development in cardiovascular imaging.

Main Methods:

• Trained a deep convolutional neural network for calcified plaque segmentation using expert-labeled CCTA slices.
• Revised and validated initial radiologist annotations for improved dataset quality.
• Compared standard ImageNet pre-training with self-supervised DINOv2 representation learning.
• Utilized Grad-CAM for visual explanation generation before and after annotation refinement.

Main Results:

• Refined annotations led to significantly improved segmentation accuracy and clearer plaque delineation.
• Grad-CAM analysis showed enhanced model attention focused on plaque and vessel regions.
• DINOv2 representations generated more spatially coherent and anatomically accurate attention maps.
• Expert radiologists qualitatively confirmed improved localization and reduced off-target activations.

Conclusions:

• High-quality, validated annotations are crucial for developing accurate and interpretable medical imaging AI.
• Representation learning impacts the reliability and clinical relevance of AI explainability outputs.
• A workflow combining annotation refinement, expert validation, and advanced features enhances AI development in cardiovascular imaging.
• Annotation quality is a critical determinant of XAI reliability, with explainability methods aiding dataset curation.