IRIS: An Intelligent Vision-Language System for Ocular Surface Diseases via Topic Tree and Scene-Driven VQA Generation

Sep 21, 2026

Uploaded by Miles

Large vision-language models show strong general capabilities, but clinical reasoning for specialized ocular surface diseases (OSDs) is limited by scarce high-fidelity multimodal instruction-tuning data. IRIS addresses this bottleneck by introducing IRIS-120K, a large OSD visual question-answering dataset, and a dual-branch generation engine that injects clinical priors. A Topic Finding Tree anchors visuals to anatomical and pathological concepts, while a scene-driven strategy synthesizes role-adaptive clinical dialogues for practical generalization. Fine-tuning a 4B-parameter VLM achieves state-of-the-art results, outperforming larger generalist and medical models up to 34B, highlighting structured knowledge injection over parameter scaling. Code, datasets, and model weights will be released.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

AccessFree to download

1 Introduction

2 IRIS and Clinically-driven Data Engine

Topic Finding Tree (TFT)

Scene-Driven Generation

arXiv :2607 .04344v 1 [ cs .CV] 5 Jul 2026 IRIS: An Intelligent Vision-Language System for Ocular Surface Diseases via Topic Tree and Scene-Driven VQA Generation Hao Wei 1⋆[0000−0002−5719−8826], Wenjin Qi2⋆[0009−0004−9816−1423], Dasen Dai 1[0009−0000−3227−3140], Minqing Zhang 1[0000−0002−7214−0569], and Wu Yuan 1[0000−0001−9405−519X]⋆⋆ 1 The Chinese University of Hong Kong, Shatin, Hong Kong, China 2 Westlake University, Xihu District, Hangzhou, Zhejiang Province, China Abstract. While Large Vision-Language Models (VLMs) demonstrate remarkable generic capabilities, their clinical reasoning in specialized domains like ocular surface diseases (OSDs) is severely hindered by a paucity of high-fidelity, multimodal instruction-tuning data. To dismantle this data bottleneck, we introduce IRIS, an Intelligent Recognition and Interaction System tailored for fine-grained OSD understanding via external eye photography. First, we curate IRIS-120K, the largest and most comprehensive OSD visual question-answering (VQA) dataset to date. Crucially, to overcome the semantic shallowness of conventional image-caption pairs, we propose a synergistic data generation paradigm to explicitly inject clinical priors. Our data engine operates via a dualbranch framework: 1) a Topic Finding Tree (TFT) that hierarchically anchors visual features to precise anatomical and pathological concepts, enforcing rigorous medical deduction logic; and 2) a Scene-driven strategy that synthesizes role-adaptive clinical dialogues to ensure pragmatic generalization. By explicitly aligning a compact 4B-parameter VLM on this structurally enriched corpus, IRIS achieves state-of-the-art performance, comprehensively outperforming both generalist and specialized medical VLMs with up to 34B parameters. Our findings underscore that structured knowledge injection profoundly prevails over sheer parameter scaling, unlocking the potential for resource-efficient, expert-level AI deployment on mobile edge devices for scalable OSD screening. Code, datasets, and model weights will be publicly released by this repo. Keywords: Vision-Language Models · Ocular Surface Diseases · Clinical Reasoning · Interpretability · Mobile Edge AI. 1 Introduction Ocular surface diseases (OSDs), including dry eye disease (DED) and infectious keratitis, represent a formidable global health challenge and the fifth leading ⋆ Equal contribution. ⋆⋆ Corresponding author ([wyuan@cuhk.edu.hk](wyuan@cuhk.edu.hk)). 2 H. Wei et al. cause of blindness [15,26] . DED afflicts up to 34% of adults, while infectious keratitis is a leading cause of blindness, which is largely preventable if diagnosed early [6,15] . Despite this prevalence, expert-level diagnosis traditionally mandates specialized stationary equipment (e.g., slit-lamp biomicroscopes) [13,26], exacerbating healthcare disparities in resource-limited settings. Consequently, the ubiquity of high-resolution smartphone cameras has catalyzed mobile eye health (MeHealth) into a transformative frontier, offering an unprecedented opportunity for accessible, large-scale OSD screening via external eye photography [23] . Ophthalmic AI has rapidly evolved from task-specific classifiers (e.g., CorneAI [14,15]) towards multimodal foundation models like RETFound [28], VisionFM [17] and EyeCLIP [19] . While these architectures demonstrate robust generalization for retinal conditions, a critical modality disparity persists: current models exhibit a profound bias towards the posterior segment. For instance, within the 2.77-million-image EyeCLIP dataset, external eye photographs account for a negligible 0.5% . Similarly, EyecareGPT [9] predominantly targets fundus and OCT imagery. This leaves a severe data and reasoning vacuum in the fine-grained understanding of external eye pathologies. While generic Large Vision-Language Models (VLMs) show potential for interactive diagnostics [21,25], recent benchmarking exposes their stark clinical limitations i

FAQs

What problem does IRIS target in ocular surface disease clinical reasoning?

It targets the lack of high-fidelity multimodal instruction-tuning data, which severely hinders accurate clinical reasoning for OSDs when using large vision-language models.

How does IRIS-120K support OSD understanding?

IRIS curates IRIS-120K as a comprehensive OSD visual question-answering dataset designed to enable fine-grained understanding from external eye photography.

What are the two key paradigms in IRIS’s data generation engine?

IRIS uses a Topic Finding Tree to hierarchically anchor visual features to ocular regions and pathological findings, and a scene-driven strategy that generates role-adaptive clinical dialogues across multiple users and scenarios.

Sep 21, 2026

Uploaded by Miles