Can Multimodal Large Language Models Understand OCT?

The paper introduces OCT-Bench, a benchmark for multimodal large language models (MLLMs) on optical coherence tomography (OCT) understanding. It contains 10,076 multiple-choice questions built from 4,137 OCT images across seven public datasets. Its taxonomy follows the clinical interpretation workflow and covers 20 fine-grained tasks across Perception, Cognition, and Reasoning, including anatomy, lesions, spatial relations, diagnosis, treatment, and prognosis. The authors evaluate 20 representative proprietary, general-purpose open-source, and medical-domain MLLMs. Results indicate that current models remain far from reliable OCT understanding, while medical adaptation and larger scale do not consistently improve performance.