Table 3.

Comparison in performance metrics of the DenseNet121-based classification model for each outcome class between the non-tilted vs. tilted disc images using the test dataset.

Outcome class	Metric	Non-tilted disc	Tilted disc	P-value
Normal	Accuracy	0.962 ± 0.008	0.945 ± 0.009	0.05
	Sensitivity	0.967 ± 0.015	0.841 ± 0.032	< 0.01
	Specificity	0.958 ± 0.013	0.976 ± 0.009	0.15
	Precision	0.940 ± 0.017	0.914 ± 0.028	0.23
	F1 score	0.953 ± 0.009	0.876 ± 0.020	< 0.01
Glaucoma	Accuracy	0.962 ± 0.009	0.939 ± 0.008	0.01
	Sensitivity	0.964 ± 0.013	0.965 ± 0.010	0.44
	Specificity	0.955 ± 0.037	0.798 ± 0.044	0.02
	Precision	0.988 ± 0.010	0.962 ± 0.008	0.03
	F1 score	0.975 ± 0.006	0.964 ± 0.005	0.06
Optic disc pallor	Accuracy	0.960 ± 0.008	0.960 ± 0.009	0.43
	Sensitivity	0.981 ± 0.012	0.980 ± 0.010	0.44
	Specificity	0.820 ± 0.058	0.608 ± 0.133	0.03
	Precision	0.974 ± 0.008	0.978 ± 0.007	0.37
	F1 score	0.977 ± 0.005	0.979 ± 0.005	0.40
Optic disc swelling	Accuracy	0.988 ± 0.005	0.985 ± 0.004	0.31
	Sensitivity	0.994 ± 0.005	0.999 ± 0.002	0.38
	Specificity	0.899 ± 0.055	0.275 ± 0.189	< 0.01
	Precision	0.993 ± 0.004	0.986 ± 0.004	0.10
	F1 score	0.994 ± 0.003	0.992 ± 0.002	0.31

The table displays the means ± standard errors of accuracy, sensitivity, specificity, precision and F1 score for the Dense121-based classification model in predicting each outcome class. P-value is a significance probability for testing the mean difference in each performance metric between classification models developed with the non-tilted vs. tilted disc images, and was derived from 100 bootstrap resamples of the test dataset.