Table 4.

Tabular data showing the quantitative performance of the various model architectures on the external test set in terms of precision, recall, and $F1$-score. Optimal performance is highlighted in bold. 95% confidence interval added for both framewise and studywise results.

	ResNet50			ResNeXt101			EfficientNetB2
Class	Precision	Recall	$F1$-score	Precision	Recall	$F1$-score	Precision	Recall	$F1$-score
Framewise external test
CA	0.829 ± 0.002	0.764 ± 0.003	0.770 ± 0.003	0.842 ± 0.002	0.798 ± 0.003	0.803 ± 0.003	0.850 ± 0.002	0.808 ± 0.003	0.813 ± 0.003
CP	0.851 ± 0.003	0.686 ± 0.004	0.739 ± 0.003	0.884 ± 0.002	0.785 ± 0.003	0.817 ± 0.003	0.859 ± 0.003	0.787 ± 0.003	0.814 ± 0.003
Normal	0.787 ± 0.004	0.800 ± 0.003	0.791 ± 0.004	0.787 ± 0.004	0.798 ± 0.003	0.791 ± 0.003	0.769 ± 0.004	0.718 ± 0.004	0.735 ± 0.003
Studywise external test
CA	0.983 ± 0.012	0.824 ± 0.027	0.896 ± 0.017	0.971 ± 0.020	0.708 ± 0.031	0.819 ± 0.022	0.982 ± 0.014	0.728 ± 0.033	0.836 ± 0.022
CP	0.207 ± 0.082	0.688 ± 0.106	0.311 ± 0.098	0.177 ± 0.075	0.752 ± 0.104	0.279 ± 0.097	0.312 ± 0.108	0.804 ± 0.097	0.440 ± 0.117
Normal	0.560 ± 0.133	0.429 ± 0.072	0.479 ± 0.078	0.580 ± 0.136	0.457 ± 0.073	0.504 ± 0.080	0.480 ± 0.132	0.682 ± 0.068	0.553 ± 0.099