Automatically Detected Number of Panretinal Photocoagulation Scars Associates with Recurrent Active Proliferative Diabetic Retinopathy: A Deep Learning Approach

Purpose
Irreversible vision-loss in diabetes is preventable if proliferative diabetic retinopathy (PDR) is diagnosed timely and treated promptly. PDR is traditionally treated with panretinal photocoagulation (PRP) but may recur despite treatment. Convolutional neural networks in deep learning represents a significant advancement in artificial intelligence, particularly capable at automated image recognition. In 2022, we developed a DR-classification model with pre-segmentation of DR-related retinal abnormalities, which performed on par with a human expert-grader in the detection of eight retinal lesions including PRP scars. We aimed to evaluate if the number of automatically detected PRP scars associates with the presence of recurrent active PDR.

Setting/venue
At a regional DR grading centre at Steno Diabetes Center Odense, Odense University Hospital, Odense, Denmark, we used the Funen Diabetes Database to identify and include six-field 45 degree retinal images of all patients with PDR that were previously treated with PRP.

Methods
We included 2,843 retinal images from patients with PDR who had previously been treated with PRP. A certified grader manually assessed all retinal images to confirm the diagnosis and stratify images according to current level of PDR (recurrent or inactive). The DL-algorithm then performed a full segmentation of PRP-lesions, which quantified the number of PRP scars.
For statistical analysis, we assessed normality using the Shapiro-Wilk test and as data were not normally distributed, we used Mann-Whitney U test to compare the number of PRP scars between eyes with recurrent and inactive PDR. We employed a univariable logistic regression analysis to investigate the association between the number of PRP scars (independent variable) and recurrent PDR (dependent variable).

Results
The certified grader detected recurrent PDR in 688 (24.2%) images and inactive PDR in 2155 (75.8%) images. The median of automatically detected PRP scars were 306 (interquartile range (IQR): 209; 400) and 239 (IQR: 158; 327) for eyes with recurrent and inactive PDR (<0.001). In a logistic regression model, the number of PRP scars associated with recurrent PDR (odds ratio 1.04 per 10 PRP scars increment, 95% CI 1.03 to 1.05).

Conclusions
In a DL-based automated detection of retinal DR-lesions, we demonstrated a higher number of PRP scars in eyes with recurrent PDR, and that the number of PRP scars served as predictor for the incidence of recurrent PDR. This highlights the necessity for ongoing regular DR screening, as sufficient PRP does not prevent recurrence of PDR. Upcoming studies should evaluate if risk of recurrent disease can also be predicted longitudinally at the time of PRP treatment.