Table 3 Performance of DLP in comparative test data set compared to experts (all fundus only, n = 922).

From: Automatic detection of 39 fundus diseases and conditions in retinal photographs using deep neural networks

Diseases/conditions	ID	Average expert				DLP
Diseases/conditions	ID	F₁	Sensitivity	Specificity	Subset accuracy^b,%	F₁	Sensitivity	Specificity	AUC (95% CI)	Subset accuracy, %
Non-referable	0	0.875	0.941	0.990		0.989	0.977	1.000	0.9955 (0.9877–1.0000)
Referable DR	1	0.951	0.935	0.993		0.970	0.953	0.997	0.9942 (0.9894–0.9989)
RVO	2	0.954	0.951	0.996		0.970	0.976	0.996	0.9984 (0.9964–1.0000)
RAO	3	0.942	0.929	0.999		1.000	1.000	1.000	1.0000 (1.0000–1.0000)
Rhegmatogenous RD	4	0.980	0.989	0.999		0.971	0.971	0.999	0.9996 (0.9991–1.0000)
Posterior serous/exudative RD	5	0.955	0.941	0.999		0.964	1.000	0.998	0.9989 (0.9976–1.0000)
Maculopathy	6	0.951	0.955	0.996		0.965	0.958	0.998	0.9967 (0.9929–1.0000)
ERM	7	0.957	0.946	0.998		0.962	0.981	0.997	0.9974 (0.9940–1.0000)
MH	8	0.957	0.945	0.999		0.939	0.939	0.998	0.9712 (0.9154–1.0000)
Pathological myopia	9	0.962	0.952	0.998		0.992	0.984	1.000	0.9999 (0.9997–1.0000)
Optic nerve degeneration	10	0.961	0.950	0.998		0.972	1.000	0.997	0.9996 (0.9988–1.0000)
Severe hypertensive retinopathy	11	0.895	0.885	0.997		0.852	1.000	0.990	0.9961 (0.9919–1.0000)
Disc swelling and elevation	12	0.963	0.955	0.998		0.970	1.000	0.997	0.9997 (0.9993–1.0000)
Dragged disc	13	0.932	0.882	1.000		0.970	0.941	1.000	0.9849 (0.9553–1.0000)
Congenital disc abnormality	14	0.860	0.891	0.998		0.952	0.909	1.000	0.9880 (0.9643–1.0000)
Pigmentary degeneration	15	0.989	0.978	1.000		1.000	1.000	1.000	1.0000 (1.0000–1.0000)
Peripheral retinal degeneration and break	16	0.966	0.948	1.000		1.000	1.000	1.000	1.0000 (1.0000–1.0000)
Myelinated nerve fiber	17	0.996	1.000	1.000		1.000	1.000	1.000	1.0000 (1.0000–1.0000)
Vitreous particles	18	1.000	1.000	1.000		0.966	1.000	0.999	1.0000 (1.0000–1.0000)
Fundus neoplasm	19	0.918	0.933	0.999		1.000	1.000	1.000	1.0000 (1.0000–1.0000)
Hard exudates	20	0.959	0.994	0.997		0.946	1.000	0.995	0.9998 (0.9994–1.0000)
Yellow-white spots/flecks	21	0.931	0.926	0.995		0.898	0.971	0.985	0.9929 (0.9868–0.9989)
Cotton-wool spots	22	0.962	0.934	0.999		0.982	0.966	1.000	0.9989 (0.9974–1.0000)
Vessel tortuosity	23	0.797	0.741	0.998		0.867	0.765	1.000	0.9917 (0.9816–1.0000)
Chorioretinal atrophy/coloboma	24	0.951	0.922	0.999		0.958	1.000	0.995	0.9980 (0.9958–1.0000)
Preretinal hemorrhage	25	0.949	0.958	0.998		0.970	0.970	0.999	0.9932 (0.9799–1.0000)
Fibrosis	26	0.947	0.935	0.998		0.976	1.000	0.998	0.9992 (0.9980–1.0000)
Laser spots	27	0.973	0.951	1.000		0.964	0.930	1.000	0.9913 (0.9803–1.0000)
Silicon oil in eye	28	0.987	0.981	1.000		1.000	1.000	1.000	1.0000 (1.0000–1.0000)
Blur fundus	29	0.833	0.750	1.000		0.750	0.750	0.999	0.9499 (0.8525–1.0000)
Referable^a, frequency-weighted average		0.954	0.943	0.998		0.964	0.972	0.998	0.9964
Total					92.45					92.19

^aReferable refers to bigclass 1–29.
^bSubset accuracy measures the scale of samples having identical labels between DLP prediction and the ground-truth labels.

Back to article page

Table 3 Performance of DLP in comparative test data set compared to experts (all fundus only, n = 922).

Search

Quick links