Automatic lesion detection in high-definition white-light bronchoscopic video

The formation of bronchial lesions along the airway walls (mucosa) is known to be a potential indicator of early lung cancer. White-light bronchoscopy (WLB) has long been the standard minimally invasive modality for identifying bronchial lesions. To do this, the physician performs a procedure known as bronchoscopy, examining the major airways to interactively locate suspect lesion sites. Unfortunately, past clinical studies have established that such an exam demands a time-consuming, careful inspection of the lengthy incoming WLB video stream. Because suspect lesions are typically distinguished from normal background regions by subtle differences, many lesions are easily missed in the video stream, resulting in lesion detection rates as low as 29% in past clinical studies. The recent introduction of high-definition (HD) WLB offers the potential for more detailed airway wall imaging, thereby offering a potentially more informative data source for identifying lesions. Nevertheless, the physician still must rely on interactive inspection to identify lesions. We consider deep learning approaches for automating the process of analyzing WLB video. In particular, we develop an HD WLB ground truth dataset and apply this dataset to a series of deep learning models to the problem of automatic bronchial lesion detection and segmentation. In our study, the best model achieved 97% detection accuracy and a Dice score of 0.54 for lesion segmentation.