Fig. 3: DeepFocus with additional score prediction.
From: DeepFocus: fast focus and astigmatism correction for electron microscopy
a Convergence of the model from Fig. 2a (using n = 10 instead of 5 patch pairs) on the image in b that contained a blood vessel; 200 ns dwell time, image size 2048 × 1568. c Model architecture that predicts an additional score si per patch pair used to calculate a weighted focus correction. d Improved convergence of the patch-score model with n = 10 input patch pairs (384 × 384). e Image after using the score model (left) and the image with correct focus (right). f One of the two input patches and corresponding score values (n = 1; ratios of maximum values; original values: 0.0065, 0.1235, 0.129). g Convergence with pixel-level score predictions using n = 2 input patch pairs (384 × 384) at 50 ns dwell time and image size 2048 × 1536. h Score map of one example patch used in g. The right column shows the composite images of the example input patch (left column) and the corresponding pixel scores (center column) in red. Scale bars: 2 µm in b,e and 0.5 µm for f and h. Data in all plots are presented as mean values ± SD from the n input patches. Source data are provided as a Source Data file.
