Near real-time nerve visualization using coherent Raman scattering rigid endoscope and deep learning-based image processing for nerve-sparing surgery

Label-free molecular imaging based on Raman scattering is attractive for medical imaging applications. The long exposure time of Raman imaging is the most significant barrier for medical applications. Here, we will present the improvement of imaging speed using deep-learning-based segmentation for coherent Raman endoscopic imaging. We used 3,600 nerve images obtained with coherent anti-Stokes Raman scattering endoscopy for the training of U-Net architecture. We investigated the shortest available exposure time relationship between the exposure time of the input images and the quality of the output images. As a result, the imaging speed accelerated to 37.5 images/min from 0.68 images/min when the segmentation quality satisfies the criterion required for medical imaging.