BrYOLO-Mamba: A Approach to Efficient Tracheal Lesion Detection in Bronchoscopy

Bronchoscopy is an important clinical tool for diagnosing and treating pulmonary diseases, playing a crucial role in the early detection of lesions and improving treatment outcomes. However, manual analysis of bronchoscopic images is both time-consuming and labor-intensive, with a higher risk of human error, particularly due to physician fatigue. Although artificial intelligence has made significant progress in enhancing the accuracy and efficiency of medical image diagnosis, current systems still struggle to effectively capture long-range dependencies and fine-grained features in complex medical images. Specifically, CNN and Transformer models face limitations in modeling long-range interactions and managing computational complexity, which reduces their effectiveness in bronchoscopic image analysis.To address these issues, this study proposes the BrYOLO-Mamba model. The model adopts the Br-ORSS Block as its core, integrating an optimized SSM into the Br-ORSS Block to fully leverage the advantages of SSMs, enabling efficient capture of both local and global dependencies in images. Additionally, improvements were made to the input and neck modules of the YOLO-Mamba architecture to enhance the detection of tracheal lesion areas. The BrYOLO-Mamba model was tested on three bronchoscopic image datasets, and the results showed significant performance improvements. Compared to the YOLO-Mamba model, BrYOLO-Mamba achieved a 2.95% increase in accuracy and a 3.41% improvement in F1 score. Moreover, BrYOLO-Mamba demonstrated high sensitivity in detecting lesion areas, further proving its excellent detection capabilities. This model not only provides a powerful solution for bronchoscopic lesion detection but also significantly reduces computational costs, offering valuable insights for future research and clinical applications.