BERD: Fine-Grained Bronchoscopy Examination Report Dataset

Bronchoscopy examination is essential for diagnosing and managing respiratory diseases, enabling the identification of abnormalities and guiding treatment decisions. Accurate bronchoscopy examination reports are crucial for effective clinical communication; however, the writing process is time-consuming and requires specialized expertise. With advancements in Multimodality Large Language Models (MLLMs), automatic report generation can reduce the workload of clinicians, improve accuracy, and enhance efficiency.However, existing bronchoscopy datasets are limited, focusing primarily on single-label classification or segmentation tasks. These datasets often describe only specific lesions, lacking the comprehensive annotations needed for the complex, multi-abnormality cases commonly seen in bronchoscopy examination. This restricts MLLMs from learning the intricate relationships between images and reports, affecting the quality of automated generation.To address this, we introduce the BERD, a Bronchoscopy Examination Report Dataset, which includes 3,697 bronchoscopy examination reports. Each report contains four selected representative images and a comprehensive description of these images. Among these images, 6,330 selected representative images are annotated with single-image text descriptions and classification labels. All these annotations are done by experienced clinicians in the bronchoscopy area. Unlike existing datasets, BERD emphasizes versatile descriptions of findings. Experimental results show that fine-tuning state-of-the-art MLLMs on our dataset significantly improves their ability to generate accurate and comprehensive reports, advancing AI applications in bronchoscopy.