Fig. 6.

Effect of image denoising on segmentation predictions. A UNet model was trained to identify noise regions (a, yellow circles) in our super-resolution images. The identified noisy regions were mapped and subtracted from the original image (a) to create a denoised version (b), in order to improve test accuracy. When denoising was not performed, noisy regions may be falsely identified as nuclei (c, yellow arrows). After denoising, most false positives due to noise disappear (d), with the occasional exception of missed noise instances (red arrow in D and red circle in B). Segmentation was conducted using Mask R-CNN