Extended Data Fig. 2: Methodology and performance of the segmentation process.
From: Three-dimensional nanoimaging of fuel cell catalyst layers
a, Schematic depiction of the training strategy as described in Methods on the LSC7 sample. For each denoised reconstruction, 1/10 z-sections are extracted and sparsely annotated by hand in 2D. This dataset is used for training a U-Net model which is then applied to predict the segmentation maps of every z-sections, that is, plane-by-plane, in the denoised volume, before reassembling in 3D. Scale cube is 203 nm3 and scale bars are 50 nm. b, e, Example tomograms and close-ups taken from the reconstructions and c, f, corresponding segmentation output for the LSC7 (b,c) and HSC7 (e,f) samples. Scale bars are 50 nm, 20 nm in the close-ups. d, g, Comparison of segmentation metric scores computed from a validation dataset held-out from training. For each class, recall is defined as the fraction of ground truth pixels of this class correctly labelled as such in the output, precision as the fraction of output pixels of this class correctly labelled. The F1 score incorporates both metrics in one.
