Single-shot optically sectioned fluorescence endomicroscopy using unsupervised RCAN-CycleGAN

Optical sectioning endomicroscopy significantly improves image contrast by rejecting out-of-focus background fluorescence, but conventional HiLo microscopy requires at least two structured-illumination exposures and, in some cases, axial scanning, which increases acquisition time and system complexity. Obtaining large, perfectly registered wide-field and HiLo image pairs for supervised deep learning is also extremely challenging in vivo. Here, we present an unsupervised RCAN-CycleGAN framework that translates single-shot wide-field fluorescence endomicroscopy images into high-contrast, optically sectioned HiLo images using only unpaired training data. Trained on unpaired wide-field and HiLo images from fluorescent beads, ex vivo mouse brain, and plant specimens, the model generates HiLo-quality output from a single wide-field exposure without structured illumination hardware. On an independent test set of paired images, our method achieves an average PSNR of 32.2 dB and SSIM of 0.9 with respect to experimental HiLo images, gains of >13 dB in PSNR and >0.7 in SSIM over raw wide-field inputs. Compared with other methods on the same dataset, the proposed RCAN-CycleGAN consistently outperforms the unsupervised baseline CUT as well as CycleGAN variants using U-Net generators. Notably, it also shows better generalization on previously unseen in vivo mouse brain data than the supervised Pix2Pix baseline. These results demonstrate that high-fidelity optical sectioning can be achieved computationally from conventional wide-field endomicroscopy in a fully unpaired manner, enabling a compact, real-time, single-shot solution that can facilitate optically sectioned imaging in clinical and resource-limited settings.