Skip to main content
. 2021 Oct 15;1:752788. doi: 10.3389/fbinf.2021.752788

FIGURE 3.

FIGURE 3

Deep learning-based density reconstruction with FRC area loss. (A) Deep learning workflow using a 2D U-Net architecture to predict density images from sparse input. (B) Example of sparse input with 10% of localizations (left), network prediction after training (middle), and target image using the full dataset (right). Scale bar = 0.5 µm (C) Total loss monitored during network training for training (black) and unseen validation images (orange). (D) FRC loss during training for target density images generated by classical isotropic Gaussian (grey) and anisotropic kernel density estimation (black). (E) Structural similarity index (MS-SSIM) during training for a network trained on FRC loss only (red), SSIM (blue) and both together (black). (F) FRC area loss during training for a network trained on FRC loss only (red), SSIM (blue), and both together (black).