Abstract:Fiber-based lensless endoscopy is powerful tool for minimally invasive tissue in clinical practice. However, the inherent honeycomb-artifact reduce the resolution and increases diagnosis difficulty. We proposed an end-to-end resolution enhancement and classification network for fiber bundle imaging. Comparing with conventional interpolation and filtering methods, the average peak signal to noise ratio (PSNR) can be improved 2~6 dB. Then we trained a VGG-19 classification network on label-free multiphoton images of 382 human braintumor 28 nontumor brain samples. The results show the classification accuracy of enhanced images is up to 91%, while the fiber bundle images are only 67% accurate. The method paves the way to in vivo histologic imaging through miniaturized endoscopic probes, and gives rapid and accurate determination for intraoperative diagnosis.

Learning-based High-Resolution Lensless Fiber Bundle Imaging for Tumor