Vascular pattern recognition for narrow-band imaging bronchoscopy

Abstract

Narrow-band imaging (NBI) is an advanced bronchoscopic technique that has shown much potential in the early detection and staging of lung cancer. NBI offers enhanced visualization of microvascular structures and mucosal morphology in the epithelium (airway walls) that are hallmarks of neoplasia. Recent studies suggest that these lesions are helpful in predicting the histological structure of the epithelial tissue in the early stages of lung cancer. Moreover, studies show a strong correlation between the invasiveness of a lesion and its vascular patterns. Shibuya first described these pathological patterns on bronchial mucosa in order to differentiate between specific histological stages of lung cancer. We propose a method to identify these vascular patterns to provide physicians with information that can later be used to sample bronchial abnormalities and perform proper histologic studies when suspecting lung cancer. We show that common deep learning based classification methods fail due to the small set of expert-labeled data available for this task. Thus, we propose a few-shot learning approach based on a Siamese network to learn and distinguish the pathological features using a small annotated dataset. Our method demonstrates improved inter-class separation in the embedding space as compared to a baseline CNN classifier model. Furthermore, we achieve an 88% overall classification accuracy on our test dataset, greatly surpassing the baseline model by 31% relative increase in accuracy.