Longitudinal Screening for Diabetic Retinopathy in a Nationwide Screening Program: Comparing Deep Learning and Human Graders

Table 3

The number of patients in the first and second screening in each cell of the contingency table (true positive, false positive, true negative, and false negative) for detecting sight threatening diabetic retinopathy by each modality.

Modality	Metric	First screening	Second screening	Difference (%)	value
DL	No. of patients	5,738	4,148	/
	True positives	669 (11.66%)	190 (4.58%)	-7.08%	<0.001
	False positives	102 (1.78%)	84 (2.03%)	+0.25%	0.3671
	True negatives	4,932 (85.95%)	3,853 (92.89%)	+6.94%	0.001
	False negatives	35 (0.61%)	21 (0.51%)	-0.1%	0.5139
HG	No. of patients	5,738	4,263	/a
	True positives	519 (9.04%)	165 (3.87%)	-5.17%	0.001
	False positives	71 (1.24%)	59 (1.38%)	+0.14%	0.5410
	True negatives	4,963 (86.49%)	3,915 (91.84%)	+5.35%	<0.001
	False negatives	185 (3.22%)	124 (2.91%)	-0.31%	0.3755

STDR: sight-threatening diabetic retinopathy; DL: deep learning; HG: human graders; value was calculated from chi-squared test for the difference between the first and second screening. value < 0.05.