BM-BronchoLC

This work presents our procedure to build a rich bronchoscopy dataset named BM-BronchoLC from 106 lung cancer and 102 non-lung cancer patients. The dataset contains annotations for anatomical landmarks and lesions, which are conducted by senior doctors of the Bach Mai hospital, Vietnam. To validate the dataset quality, we investigate two typical AI backbone models for the image segmentation and classification task, i.e., Unet++ and ESFPNet, using both unitask and multitask settings. Preliminary results show the significant potential of the data in building AI models or real-time diagnostic accuracy. Further research will validate these promising developments and explore broader clinical applications in this evolving field.

本研究详述了构建名为BM-BronchoLC的丰富完备支气管镜检查数据集的完整流程，该数据集源自106例肺癌患者与102例非肺癌患者。该数据集包含解剖标志点与病变区域的标注信息，标注工作由越南巴克迈医院（Bach Mai Hospital）的资深医师完成。为验证该数据集的质量，本研究采用Unet++与ESFPNet两款典型的AI主干模型，分别在单任务与多任务设置下，针对图像分割与分类任务开展实验验证。初步实验结果显示，该数据集在AI模型构建以及实时诊断精度提升领域具备显著应用潜力。后续研究将进一步验证这些颇具前景的研究成果，并在这一快速发展的领域中探索更广泛的临床应用场景。