An Explainable Multimodal Framework with LLM Agents for Intracranial Hemorrhage Detection

Abstract

Explainability in intracranial hemorrhage (ICH) diagnosis is essential for timely and accurate clinical decisions, especially in life–threatening situations. We propose a framework that generates explainable, clinically relevant text from 2D CT scans using two cooperative GPT-4o agents: a Multi-modal User Agent (MUA) and a Planner Agent. The MUA interprets scans with YOLOv10 (mosaic augmentation), SAM2, and clustering

the Planner selects tools and outputs key imaging parameters: bleed location, midline shift, calvarial fracture, and mass effect crucial for urgent interventions. Explainability is enforced via chain-of-thought prompting to ensure transparent decision-making. Experiments show YOLOv10 with mosaic improves mAP@0.5:0.95 by 4.1% over existing methods, and the LLM agents extract clinical parameters with 78.1% accuracy (Our code is available at https://github.com/Shashwathp/Explainable-ICH-Detection-with-LLM-Agents/tree/main). These results underscore the potential of explainable AI to enhance trust and reliability in critical healthcare applications. © 2025 Elsevier B.V., All rights reserved.