Table 2.

Diagnostic performance of deep learning models in individual studies

Study	RD subtype	Test set (RD/non-RD)	TP	FN	FP	TN	Sensitivity	Specificity	AUC
Ohsugi 2017 [26]	RRD	411/420	401	10	15	405	0.98	0.96	0.988
Li 2020 [19]	Mixed-type RD	200/200	199	1	1	199	1.00	1.00	1.000
Zhang 2021 [20]	Mixed-type RD	16/173	14	2	0	173	0.88	1.00	1.000
Zhou 2022 [21]	Recurrent/postoperative RD	54/35	45	9	4	31	0.83	0.89	0.930
Cao 2022 [22]	Mixed-type RD	532/3162	529	3	7	3155	0.99	1	NR
Sun 2023 [23]	Mixed-type RD	135/844	132	3	8	836	0.98	0.99	0.998
Antaki 2023 [13]	Mixed-type RD	88/125	70	18	4	121	0.80	0.97	NR
Tang 2023 [24]	RRD	146/735	134	12	8	727	0.92	0.99	0.988
Cui 2023 [25]	Mixed-type RD	6/6216	3	3	31	6185	0.50	1.00	0.910
Wang 2023 [14]	RRD	65/535	57	8	7	528	0.88	0.99	NR
Christ 2024 [27]	Mixed-type RD	176/305	153	23	14	291	0.87	0.95	0.972

Study-level diagnostic performance of deep learning models for detecting retinal detachment. “Test set (RD/non-RD)” indicates the number of retinal detachment and non-retinal detachment images (or eyes) in the primary test set selected for each study. Sensitivity was calculated as TP/(TP + FN) and specificity as TN/(TN + FP). AUC values are those reported by the original authors for the corresponding binary RD versus non-RD task; NR indicates that AUC was not reported. For studies that did not report TP, FN, FP, and TN, these values were back-calculated from the reported sensitivity, specificity, and test-set sizes. “Mixed-type RD” indicates cohorts in which the RD group included RRD and other RD subtypes; cohorts restricted to recurrent/postoperative RD were treated as an extreme-spectrum population

Abbreviations: RD = retinal detachment; RRD = rhegmatogenous retinal detachment; TP = true positive; FN = false negative; FP = false positive; TN = true negative; AUC = area under the receiver operating characteristic curve; NR = not reported