Comparison of Autonomous AS-OCT Deep Learning Algorithm and Clinical Dry Eye Tests in Diagnosis of Dry Eye Disease

Figures & data

Figure 1 Anterior segment optical coherence tomography of a healthy cornea using an HD-OCT (Envisu R2210, Bioptigen, Leica, Buffalo Grove, IL, USA) collected at the Bascom Palmer Eye Institute.

Figure 2 Flowchart of the distribution of images throughout the training and testing of the deep learning model, including the quality control phase and grouping of images within the healthy and dry eye disease groups.

Table 1 Demographics

	Healthy	Dry Eye Syndrome	Total
Images, no.	3060	3960	7020
Patients, no.	10	14	24
Eyes, no. (%)	17	22	39
Right	8 (47.1%)	12 (54.5%)	20 (51.3%)
Left	9 (52.9%)	10 (45.5%)	19 (48.7%)
Age (years), mean (SD)	41.6 (15.4)	68.2 (8.3)	55.3 (18.1)
Gender, no. (%)
Male	6 (60.0%)	3 (20.0%)	9 (37.5%)
Female	4 (40.0%)	11 (80.0%)	15 (62.5%)

Note: Data from the demographics of patients collected for the testing phase after the creation of the deep learning model.

Figure 3 Training accuracy, validation accuracy, and cross entropy loss function plots during training phase of deep learning model. Training accuracy (performance in percentage) to correctly identify a trained image, validation accuracy (performance in percentage) to correctly identify a nontrained image (A) and cross entropy loss function (B) over 5710 iterations or training steps.

Figure 4 Results of the occlusion testing showing features of the OCT image recognized by the deep learning model. (A) AS-OCT images of a cornea with dry eye disease on the left and a healthy cornea on the right with (B) visualization of the learned features of the deep learning neural network.

Table 2 Results from the Testing Phase

			Corneal Fluorescein Staining		Conjunctival Lissamine Green Staining		OSDI		TBUT		Schirmer’s Test		Total
			Healthy	Dry Eye	Healthy	Dry Eye	Healthy	Dry Eye	Healthy	Dry Eye	Healthy	Dry Eye
Deep Learning Model	Healthy	Count	3	0	2	1	0	3	1	2	1	2	3
		% of Total	13.64%	0.0%	9.09%	4.55%	0%	13.64%	4.55%	9.09%	4.55%	9.09%	13.64%
	Dry Eye	Count	12	7	15	4	8	11	7	12	11	8	19
		% of Total	54.55%	31.82%	68.18%	18.18%	36.36%	50%	31.82%	54.55%	50%	36.36%	86.36%
Total		Count	15	7	17	5	8	14	8	14	12	10	22
		% of Total	68.18%	31.82%	77.27%	22.73%	36.36%	63.64%	36.36%	63.64%	54.55%	45.45%	100.0%

Notes: Data showing the percentage identified correctly, using diagnosis by a trained masked ophthalmologist as the gold standard, by the deep learning model compared independently with the five clinical dry eye tests – corneal fluorescein staining, conjunctival lissamine green staining, Ocular Surface Disease Index (OSDI), tear break-up time (TBUT), and Schirmer’s test. The deep learning model agreed with the ophthalmologist significantly more than the corneal fluorescein staining, conjunctival lissamine green staining, and the Schirmer's test (P < 0.05). There was no statistical significance between the deep learning model and the OSDI and TBUT.