Table 2 Quantification of self-regulation on in-domain and out-of-domain datasets

From: Accelerating domain-aware electron microscopy analysis using deep learning models with synthetic data and image-wide confidence scoring

Image inclusivity

High

Standard

Low

Individual prediction & image confidence score threshold

0.45 & 0.65

0.4 & 0.7

0.35 & 0.75

Inference det

Image count

Image filtering rate with varied thresholds

Cavity Test Set

20

0%

25%

35%

Expanded Cavity Set

175

16%

41%

60%

Loops Set28

107

24%

60%

91%

COCO 12831

128

72%

88%

95%

  1. The highly sensitive nature of both filtering thresholds is apparent in the significant changes in filtering rates with only minor threshold adjustments. The same Synthetic model was used in all evaluations.
Back to article page