Table 8. Classification results (in percentage) at the B-scan level on UCSD dataset.

Methods	Classes	ACC	SE	PR	SP	AUC	OA	OS	OP
LACNN [24]	CNV	92.7 ± 1.5	89.8 ± 4.5	93.5 ± 1.3	95.1 ± 1.6	97.7 ± 0.5	90.1 ± 1.4	86.8 ± 1.3	86.2 ± 2.3
	DME	96.6 ± 0.2	87.5 ± 1.5	86.4 ± 1.6	98.0 ± 0.3	97.4 ± 0.4
	DRUSEN	93.6 ± 1.4	72.5 ± 7.9	70.0 ± 5.7	95.9 ± 2.1	93.4 ± 1.5
	NORMAL	97.4 ± 0.2	97.3 ± 1.0	94.8 ± 1.1	97.4 ± 0.5	99.2 ± 0.2
Our model (with the pre-trained weights on our development OCT dataset)	CNV	92.3 ± 1.2	88.4 ± 3.1	93.7 ± 1.5	95.7 ± 1.5	97. 2 ± 0.4	90.4 ± 1.2	87.2 ± 1.4	86.7 ± 1.8
	DME	97.1 ± 0.2	90.1 ± 1.2	94.3 ± 1.3	98.2 ± 0.4	97.8 ± 0.3
	DRUSEN	93.7 ± 1.3	73.1 ± 6.5	71.2 ± 4.3	96.3 ± 1.8	93.9 ± 1.1
	NORMAL	98.3 ± 0.2	98.1 ± 0.7	96.2 ± 1.0	98.4 ± 0.3	99.8 ± 0.08
Our model (without using the pre-trained weights)	CNV	92.1 ± 1.3	88.1 ± 3.3	93.4 ± 1.6	95.5 ± 1.6	97. 1 ± 0.5	90.2 ± 1.3	86.9 ± 1.5	86.5 ± 2.0
	DME	96.9 ± 0.2	89.2 ± 1.3	93.1 ± 1.4	98.1 ± 0.5	97.7 ± 0.4
	DRUSEN	93.5 ± 1.5	72.8 ± 7.1	70..7 ± 4.9	96.1 ± 1.9	93.7 ± 1.2
	NORMAL	97.9 ± 0.2	97.7 ± 0.8	95.6 ± 1.1	98.0 ± 0.4	99.5 ± 0.1

UCSD: University of California San Diego; ACC: accuracy; SE: sensitivity; PR: precision; SP: specificity; AUC: area under receiver operating characteristic curve; OA: overall accuracy; OS: overall sensitivity; OP: overall precision; LACNN: lesion-aware convolutional neural network; CNV: choroidal neovascularization; DME: diabetic macular edema; OCT: optical coherence tomography.