Table 2 RetinaNet and YOLOv5 single-model results comparing performance with the standard fixed decision threshold and applying the various avalanche schemes. \(\gamma\) represents the constant rate reduction between each decision threshold in the avalanche scheme.
From: High sensitivity methods for automated rib fracture detection in pediatric radiographs
Model | Scheme | Precision | Recall | F2 |
|---|---|---|---|---|
RetinaNet | Standard | 0.892 ± 0.015 | 0.430 ± 0.014 | 0.480 ± 0.014 |
Posterior | 0.141 ± 0.015 | 0.872 ± 0.013 | 0.427 ± 0.026 | |
Conservative | 0.530 ± 0.023 | 0.730 ± 0.015 | 0.679 ± 0.010 | |
\(\gamma =0.15\) | 0.304 ± 0.028 | 0.766 ± 0.015 | 0.586 ± 0.019 | |
\(\gamma =0.20\) | 0.256 ± 0.023 | 0.770 ± 0.024 | 0.548 ± 0.015 | |
YOLOv5 | Standard | 0.897 ± 0.032 | 0.434 ± 0.040 | 0.484 ± 0.037 |
Posterior | 0.759 ± 0.164 | 0.647 ± 0.101 | 0.652 ± 0.051 | |
Conservative | 0.831 ± 0.101 | 0.590 ± 0.075 | 0.622 ± 0.053 | |
\(\gamma =0.15\) | 0.814 ± 0.120 | 0.587 ± 0.077 | 0.615 ± 0.050 | |
\(\gamma =0.20\) | 0.816 ± 0.114 | 0.593 ± 0.087 | 0.621 ± 0.060 |
