Table 1 Comparison against common model architectures using the mean absolute error (MAE)

From: DeepFocus: fast focus and astigmatism correction for electron microscopy

Architecture

MAEwd [µm]

MAEstigx [a.u.]

MAEstigy [a.u.]

GPU speed [MPx/s]

CPU speed [MPx/s]

Stacked Conv.

2.59 ± 2.39

0.45 ± 0.39

0.43 ± 0.30

518.26 ± 146.90

54.00 ± 0.74

ResNet-50

3.06 ± 4.24

0.52 ± 0.56

0.43 ± 0.42

102.76 ± 1.24

2.60 ± 0.02

EfficientNet-B0

1.45 ± 1.43

0.29 ± 0.21

0.37 ± 0.33

406.56 ± 22.25

14.14 ± 0.28

  1. Single iteration and single input patch-pair performance (mean ± SD) of different model architectures (Stacked Conv. with an edge length of 512) on the neural tissue test data and pixel throughput (Methods).
Back to article page