Table 2.

LightOCT model performance on the AIIMS¹⁴, Srinivasan’s¹⁶ and Kermany’s¹⁵ datasets with training, validation and testing sets split using different strategies.

Dataset	Split strategy	MCC [−1, 1] (m ± std)	AUC [0,1] (m ± std)	F1-score [0,1] (m ± std)	Accuracy [0,1] (m ± std)	Precision [0,1] (m ± std)	Recall [0,1] (m ± std)
AIIMS14	per-image	0.958 ± 0.038	1.000 ± 0.000	0.978 ± 0.021	0.978 ± 0.021	1.000 ± 0.000	0.978 ± 0.021
	per-volume/subject	0.881 ± 0.102	0.996 ± 0.009	0.934 ± 0.063	0.935 ± 0.060	0.993 ± 0.014	0.935 ± 0.060
Srinivasan16	per-image	0.853 ± 0.039	0.985 ± 0.005	0.898 ± 0.030	0.899 ± 0.029	0.973 ± 0.009	0.899 ± 0.029
	per-volume/subject	0.426 ± 0.116	0.817 ± 0.055	0.593 ± 0.088	0.603 ± 0.078	0.702 ± 0.078	0.603 ± 0.078
Kermany15 version 2	original_v2	0.886	0.993	0.909	0.911	0.983	0.911
	per-image	0.707 ± 0.021	0.953 ± 0.003	0.764 ± 0.022	0.770 ± 0.019	0.886 ± 0.007	0.770 ± 0.019
	per-volume/subject	0.588 ± 0.025	0.890 ± 0.006	0.644 ± 0.033	0.669 ± 0.023	0.769 ± 0.012	0.669 ± 0.023
Kermany15 version 3	original_v3	0.644	0.964	0.678	0.704	0.916	0.704
	per-image	0.673 ± 0.021	0.950 ± 0.003	0.729 ± 0.022	0.738 ± 0.019	0.886 ± 0.007	0.738 ± 0.019
	per-volume/subject	0.600 ± 0.021	0.911 ± 0.006	0.651 ± 0.028	0.671 ± 0.021	0.795 ± 0.012	0.671 ± 0.021

Performance metrics are reported as mean ± standard deviation (m ± std) over the models trained through ten-times repeated five-fold cross validation and classes, for the per-image and per-volume/subject splits. For the original splits given by Kermany, results are reported for the single given split. AUC: area under the receiver operating characteristic curve, MCC: Matthews Correlation Coefficient.