Prediction of Causative Genes in Inherited Retinal Disorders from Spectral-Domain Optical Coherence Tomography Utilizing Deep Learning Techniques

Table 1

Summary of deep learning performance in prediction of causative genes in inherited retinal disorders.

Experiment 1.		Training results				Test 1.		Test results						Total number of images included in this study
		Number of images	Sensitivity (%)	Specificity (%)	Accuracy (%)			Original category of genetic diagnosis					Accuracy (%)
								ABCA4	RP1L1	EYS	Normal	Total
Original classification of genetic diagnosis	ABCA4	15	100	100	—	Predicted category of genetic diagnosis	ABCA4	4	1	—	—	5	100	ABCA4	19
	RP1L1	29	85.7	100	—		RP1L1	—	7	—	—	7	87.5	RP1L1	37
	EYS	43	100	95	—		EYS	—	—	14	—	14	100	EYS	57
	Normal	49	100	100	—		Normal	—	—	—	16	16	100	Normal	65
	Total	136	—	—	96.9		Total	4	8	14	16	42	97.6	Total	178
Experiment 2.		Training results				Test 2.		Test results						Total number of images included in this study
		Number of images	Sensitivity (%)	Specificity (%)	Accuracy (%)			Original category of genetic diagnosis					Accuracy (%)
								ABCA4	RP1L1	EYS	Normal	Total
Original classification of genetic diagnosis	ABCA4	13	100	100	—	Predicted category of genetic diagnosis	ABCA4	6	2	1	—	9	100	ABCA4	19
	RP1L1	26	100	100	—		RP1L1	—	9	1	2	12	82	RP1L1	37
	EYS	43	100	100	—		EYS	—	—	12	3	15	85.7	EYS	57
	Normal	46	100	100	—		Normal	—	—	——	14	14	73.7	Normal	65
	Total	128	—	—	100		Total	6	11	14	19	50	82	Total	178
Experiment 3.		Training results				Test 3.		Test results						Total number of images included in this study
		Number of images	Sensitivity (%)	Specificity (%)	Accuracy (%)			Original category of genetic diagnosis					Accuracy (%)
								ABCA4	RP1L1	EYS	Normal	Total
Original classification of genetic diagnosis	ABCA4	15	100	96.6	—	Predicted category of genetic diagnosis	ABCA4	4	—	1	—	5	100	ABCA4	19
	RP1L1	31	66.7	100	—		RP1L1	—	4	—	—	4	66.7	RP1L1	37
	EYS	45	90.9	90.5	—		EYS	—	2	10	—	12	90.9	EYS	57
	Normal	49	100	100	—		Normal	—	—	—	16	16	100	Normal	65
	Total	140	—	—	90.6		Total	4	6	11	16	37	91.9	Total	178
Experiment 4.		Training results				Test 4.		Test results					Accuracy (%)	Total number of images included in this study
		Number of images	Sensitivity (%)	Specificity (%)	Accuracy (%)			Original category of genetic diagnosis
								ABCA4	RP1L1	EYS	Normal	Total
Original classification of genetic diagnosis	ABCA4	14	100	100	—	Predicted category of genetic diagnosis	ABCA4	5	—	—	—	5	100	ABCA4	19
	RP1L1	25	100	100	—		RP1L1	—	10	1	—	11	76.9	RP1L1	37
	EYS	40	100	100	—		EYS	—	—	14	—	14	82.4	EYS	57
	Normal	51	100	100	—		Normal	—	2	2	14	18	100	Normal	65
	Total	130	—	—	100		Total	5	13	17	14	49	89.8	Total	178

In total, 75 subjects with molecularly confirmed inherited retinal disorders or no ocular diseases have been ascertained: 10 with ABCA4 retinopathy, 20 patients with RP1L1 retinopathy, 28 with EYS retinopathy, and 17 normal subjects. After preparation of spectral-domain optical coherence tomographic (SD-OCT) images for four gene categories, subjects were randomly split following a 3 : 1 ratio into training and test sets. The commercially available deep learning web tool, Medic Mind, was applied to this four-class classification problem. The classification accuracy, sensitivity, and specificity were calculated during the learning process, and the process was repeated four times with randomly assigned training/test sets to control for selection bias. For each training/testing process, the classification accuracy was calculated per gene category.