Abstract
Multi-modal large language models (MLLMs) show promise for computer-aided medical diagnosis, yet their deployment in ophthalmology still lacks interpretability and trustworthiness. While integrating specialized expert tools provides essential fine-grained perception, it introduces a critical bottleneck: cognitive offloading. Current agentic MLLMs often degenerate into passive execution engines, blindly trusting brittle, domain-specific tools that are highly susceptible to clinical noise and image artifacts. Consequently, they forfeit their intrinsic critical thinking and suffer from error cascades under complex fundus manifestations. To address this paradigm flaw, we introduce OphthaAgent, a versatile tool-augmented agentic MLLM for fundus diagnosis that enables fine-grained, evidence-based reasoning through an iterative evidence acquisition and verification mechanism. Operating in an online, multi-turn interactive environment, OphthaAgent is trained to dynamically invoke specialist tools and trigger reflective verification based on real-time observations. Extensive experiments on two fundus benchmarks demonstrate state-of-the-art performance in both in-distribution and out-of-distribution settings and further reveal an emergent self-correction ability in OphthaAgent. Project page: https://github.com/rikey123/OphthaAgent.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/OMIA_007.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=yMbyDvIJ8p
BibTex
@InProceedings{XieWen_OphthaAgent_MICCAISAT2026,
author = { Xie, Wenqing AND Wang, Ruiqi AND Dai, Gaole AND Zhou, Tao AND Zhou, Yi},
title = { { OphthaAgent: Learning Cognitive Prudence for Reliable Fundus Diagnosis via Tool-Augmented Reinforcement Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17270},
month = {pending},
page = {pending}
}
