Segmentation And Quantification Of Cellular Load In Pulmonary Endomicroscopic Images Using Convolutional Neural Networks

Fibre Bundle Endomicroscopy (FBEμ) is an emerging tool that facilitates the real-time structural and functional (via fluorescent dyes) imaging of the distal lung, providing valuable in vivo, in situ indicators across a range pathological or physiological processes. This paper proposes a novel approach for localising and quantifying abnormalities in distal lung, such as increased cellular load, through semantic image segmentation. Two Convolutional Neural Network (CNN) architectures have been tested, (i) U-Net, a purpose specific network for biomedical image applications, and (ii) ENet, a network optimised for fast inference. The results indicate that semantic segmentation of cells as well as quantification of cellular load is viable, with U-Net consistently outperforming ENet, obtaining a pixel accuracy of 0.842 and a correlation (r) with the corresponding manual cellular load estimation of 0. 866.