PlaqueCap: lesion-centered captioning of atherosclerotic plaques in intravascular ultrasound using vision-language models and prompt injection.

Accurate characterization of atherosclerotic plaques in intravascular ultrasound (IVUS) imaging is essential for evaluating coronary artery disease and guiding clinical interventions. Traditional methods rely on handcrafted features and rule-based algorithms, which lack adaptability to diverse lesion morphologies and offer limited explainability. To address these challenges, this work introduces PlaqueCap, a lesion-centered captioning framework that generates clinically meaningful, natural language descriptions directly from IVUS images. A central challenge is ensuring the generated text is grounded in the specific pathology of the lesion. PlaqueCap solves this by performing high-fidelity segmentation to localize the plaque, then using a Lesion Prompt Injection (LPI) module to inject spatial information into a pre-trained vision-language model, focusing on pathological characteristics. Experimental results on a curated IVUS dataset show PlaqueCap achieves accurate lesion localization and classification, producing detailed, clinically interpretable descriptions surpassing baselines in quantitative metrics and expert evaluation. This offers a paradigm for explainable AI in intravascular imaging and automated reporting in interventional cardiology.