Generation of comparative thrombosis reports from confocal laser‑scanning microscope images using large language models

This paper presents an automated framework for generating thrombus analysis reports from confocal laser scanning microscope (CLSM) images. CLSM is widely used to image the mouse cerebral cortex to investigate disease mechanisms and thrombus formation. Generating clear and structured experimental reports that summarize current observations and inform subsequent experiments is essential but time-consuming. Prior work extracted thrombus-related features from paired images and used a large language model (LLM) to produce reports; however, without task-specific constraints, the generated text can be verbose and poorly organized. We propose an LLM-based method that generates structured findings by comparing paired CLSM images. An encoder–decoder segmentation network is first used to segment BVs and thrombi in each image. A feature fusion module is used to extract feature from color images and mask images and then integrate them with a tokenized prompt to construct the input representation for the LLM. The LLM outputs the final report text. Experiments showed that the proposed method generates accurate, well-structured, task-specific reports, improving both reporting quality and experimental throughput in biomedical research.