Background: Vascular anomalies (VAs), including hemangiomas and vascular malformations, present a significant diagnostic challenge due to their high prevalence, complex classification (nearly 100 subtypes), and visual mimicry. Current Multimodal Large Language Models (MLLMs) struggle in this specialized domain, often failing to capture fine-grained visual features or lacking evidence-based reasoning. To address these limitations, we introduce HevaDx, an agentic diagnostic system that explicitly decouples visual perception from clinical reasoning. Methods: Leveraging a newly constructed large-scale dataset of VA patients, HevaDx employs a lightweight visual specialist for precise feature extraction and a reasoning specialist equipped with Retrieval-Augmented Generation (RAG) for therapeutic planning. This cooperative architecture mitigates the “reasoning gap” observed in end-to-end models by grounding decisions in up-to-date clinical guidelines. Results: Experimental results demonstrate that HevaDx achieves high performance with a top-3 diagnostic accuracy of 94.8% and a treatment recommendation accuracy of 83.3%. Conclusions: By bridging visual precision with transparent, verifiable logic, HevaDx offers a reliable framework for AI-assisted management of vascular anomalies.
Zhang et al. (Fri,) studied this question.