XMedFusion: multi-agent AI boosts radiology report accuracy by 33%

XMedFusion addresses a core weakness in end-to-end multimodal models for radiology report generation: weak visual grounding that leads to omissions and unreliable interpretations. Developed by Riaz et al., the framework mimics expert diagnostic workflows by splitting the task into four coordinated agents. A visual perception agent extracts image-grounded evidence, a knowledge graph agent structures clinically relevant findings, a retrieval-guided drafting agent ensures consistent report structure, and a synthesis agent iteratively integrates visual and structured evidence through reasoning-driven verification. This modular design provides transparency and robustness, making it suitable for integration into autonomous healthcare and robotic diagnostic workflows.

On a public chest radiograph dataset, XMedFusion demonstrated substantial gains over baseline vision-language models: BLEU-1 improved by up to 0.3359, ROUGE-L by up to 0.2440, and METEOR by up to 0.1708. More critically, semantic metrics jumped—Consistency from 2.38 to 7.80 and Accuracy from 2.34 to 6.93—indicating that the framework not only matches reference texts better but also produces clinically coherent and correct findings. The paper, accepted at ICRAI 2026, highlights how structured multi-agent reasoning can enhance automation, trust, and reliability in medical imaging AI.