Table 4 Comparison of diagnostic performance between CNN and radiologists in previous studies.

From: Diagnosis of thyroid micronodules on ultrasound using a deep convolutional neural network

Author

Training set

Internal test set

External test set

Performances

Wang et al.25

5007 nodules

351 nodules, including 151 nodules < 1 cm

N/A

CNN showed significantly higher specificity and AUC than radiologists with comparable sensitivity

In the subgroup of nodules < 1 cm, CNN also showed significantly higher specificity than radiologists

Li et al.19

42,952 patients

1118 patients

1574 patients

CNN showed significantly lower sensitivity and higher specificity than radiologists in both internal and external test sets

Buda et al.22

1278 nodules

99 nodules

N/A

CNN showed significantly higher specificity than inexperienced radiologists who did not use ACR TIRADS

CNN showed similar AUC, sensitivity, and specificity to expert radiologists on the ACR TIRADS committee

Kim et al.24

Commercially available CAD

218 nodules

≥ 5 mm

N/A

CNN showed significantly lower specificity and AUC than radiologists with comparable sensitivity

Park et al.20

4919 nodules

≥ 5 mm

286 nodules

≥ 5 mm

N/A

No significant difference in diagnostic performance between the CNN and radiologists

Ko et al.18

439 nodules

≥ 1 cm and < 2 cm

150 nodules

≥ 1 cm and < 2 cm

N/A

No significant difference in diagnostic performance between the CNN and radiologists

Koh et al.21

13,560 nodules

≥ 1 cm

200 nodules

≥ 1 cm

600 nodules

≥ 1 cm

CNN showed significantly higher AUC in the internal test set, while no significant difference was shown in the external test sets

CNN showed significantly lower sensitivity and higher specificity than radiologists in the internal test set and one of the four external test sets

Han et al.23

Commercially available CAD

454 nodules

≥ 1 cm

N/A

CNN showed significantly lower specificity and AUC than radiologists with comparable sensitivity

  1. CNN convolutional neural network, N/A not applicable, AUC area under the receiver operating characteristics curve, ACR American College of Radiology, TIRADS Thyroid Imaging Reporting and Data System, CAD computer-aided diagnosis.