Deep Learning Assisted-Axial Resolution Improvement for Probe-Based Confocal Laser Endomicroscopy.

Abstract

Probe-based confocal laser endomicroscopy (pCLE), as a powerful medical device for detecting digestive tract diseases, can directly conduct real-time and high-resolution histological diagnosis of gastric mucosal epithelium and subcutaneous tissues. It can achieve optical biopsy and has great potential for early diagnosis of gastric cancer. Deep tissue detection and high axial resolution in the depth direction are of great significance for the diagnosis of clinical gastrointestinal diseases. However, improvements in hardware systems will greatly increase the complexity of the system structure, thereby limiting its wide application. Based on the residual channel attention network (RCAN), we developed a deep learning method for axial resolution enhancement. This method, abbreviated as ARE-RCAN, can directly reconstruct images captured by pCLE into ones with higher axial resolution without any additional hardware design. Firstly, high-quality training and testing datasets are generated through numerical simulation, and based on this, the ARE-RCAN is trained and performance tested. The reconstructed results of the testing datasets show that the axial resolution of the reconstructed images of simulated fluorescent beads has doubled, and the lateral resolution has not changed. Finally, a home-built pCLE system was used to collect fluorescent bead and paper fibers data, and the effectiveness of the ARE-RCAN was verified through experiments. The results indicate that the ARE-RCAN method significantly improves the axial resolution of the actual system.