OCT-Bench sets 10,076 questions to test whether multimodal models really understand retinal scans

Baochen Fu · hf · 2026-07-21

OCT-Bench introduces a 10,076-question benchmark for multimodal model understanding of OCT scans

The paper proposes OCT-Bench, a new benchmark for evaluating whether multimodal large language models can truly understand optical coherence tomography (OCT) images beyond coarse disease classification.

Scale: 10,076 multiple-choice questions built from 4,137 OCT images across 7 public datasets.
Task design: a 20-task hierarchy spanning three levels — Perception, Cognition, and Reasoning.
Coverage: imaging attributes, retinal anatomy, lesion characteristics, spatial relationships, disease assessment, treatment decisions, and prognosis management.
Evaluation: 20 representative MLLMs were tested, including proprietary, open-source general-purpose, and medical-domain models.
Result: current models are still far from reliable OCT understanding; neither medical adaptation nor larger scale consistently improves performance across capability levels.

The benchmark is positioned as a more clinically grounded way to identify where multimodal models fail in medical image understanding.

Original post →

OCT-Bench sets 10,076 questions to test whether multimodal models really understand retinal scans

More from Multimodal