Figure 4
At training time the loss function depends on denoising two OCT B-scans. So the same U-Net is used to separately denoise each of the two images. In contrast to that, at inference time, the loss function does not need to be computed anymore. Consequently, since the network in the training phase is used separately on a single OCT B-scan, a single OCT B-scan can be denoised on its own in this phase.
