Table 3 Figure of merit (FOM) (95% confidence interval) of jackknife free-response receiver operating curve (JAFROC) and reproducibility comparisons in terms of PPA and CPPA at five-different cutoff values (the number of false positives per CXR was 0.1, 0.2, 0.3, 0.4, and 0.5) of four CNNs based detection algorithms and two readers on nodule and mass case.

From: Short-term Reproducibility of Pulmonary Nodule and Mass Detection in Chest Radiographs: Comparison among Radiologists and Four Different Computer-Aided Detections with Convolutional Neural Net

Algorithms and readers

FOM (95% CI)

PPA (%)

CPPA (%)

Mask R-CNN

0.87 (0.83–0.91)

96.52 ± 0.51

93.18 ± 0.81

RetinaNet

0.84 (0.78–0.88)*

94.23 ± 0.00

89.09 ± 0.00

eDenseYOLO

0.89 (0.85–0.93)

97.87 ± 0.08

95.80 ± 0.22

ASPP U-Net

0.85 (0.80–0.89)*

95.04 ± 0.11

90.57 ± 0.19

R1

 

96.55

93.33

R2

—

94.98

90.43

  1. (*p-value < 0.05 between eDenseYOLO and others for FOM).
Back to article page