Deep-learning-assisted disposable bronchoscopy with enhanced image quality

The COVID-19 pandemic underscored the need for safer and more cost-effective bronchoscopy solutions, leading to increased interest in single-use bronchoscopes due to their elimination of cross-contamination risk. However, their widespread adoption has been limited by inferior image quality compared to the reusable counterparts. Addressing this need, we present in this study a deep-learning framework specifically designed to enhance the imaging performance of single-use bronchoscopes, improving the signal-to-noise ratio and image resolution while maintaining real-time processing capability. Our approach integrates a generative adversarial network architecture with attention mechanisms that focus on airway assessment. We also introduce a combined loss function to close the quality gap between reusable and disposable bronchoscopes. Additionally, we propose Fourier spectra comparison as a performance metric to evaluate the neural network performance on improving real bronchoscopy videos in the absence of ground truth images. The neural network model was initially trained and evaluated on a simulated dataset, achieving an 89% improvement in PSNR and a 16% increase in SSIM. When applied to real bronchoscopy videos, the trained model demonstrated a 1.8-fold increase in resolution based on the Fourier-spectra metric. The reconstruction speed was also computed for images of varying sizes, confirming the network's feasibility for real-time applications with large fields of view. This work not only presents a scalable AI solution that transforms disposable bronchoscopes into competitive alternatives for patients with respiratory disease but also establishes new methodological directions for evaluating and optimizing highly specialized medical imaging tasks in resource-constrained settings.