Abstract
Confocal laser endomicroscopy (CLE) is a novel imaging modality that providesin vivo histological cross-sections of examined tissue. Recently, attempts havebeen made to develop miniaturized in vivo imaging devices, specificallyconfocal laser microscopes, for both clinical and research applications.However, current implementations of miniature CLE components, such as confocallenses, compromise image resolution, signal-to-noise ratio, or both, whichnegatively impacts the utility of in vivo imaging. In this work, we demonstratethat software-based techniques can be used to recover lost information due toendomicroscopy hardware miniaturization and reconstruct images of higherresolution. Particularly, a densely connected convolutional neural network isused to reconstruct a high-resolution CLE image from a low-resolution input. Inthe proposed network, each layer is directly connected to all subsequentlayers, which results in an effective combination of low-level and high-levelfeatures and efficient information flow throughout the network. To train andevaluate our network, we use a dataset of 181 high-resolution CLE images. Bothquantitative and qualitative results indicate superiority of the proposednetwork compared to traditional interpolation techniques and competinglearning-based methods. This work demonstrates that software-basedsuper-resolution is a viable approach to compensate for loss of resolution dueto endoscopic hardware miniaturization.