Background: Diabetic Macular Edema (DME) is a complication of diabetes which, when untreated, leads to vision loss. Screening for DME is recommended for diabetic patients every 1-2 years, however compliance rates are low. Though there is currently no high-efficacy camera-agnostic system for DME detection, an AI system may improve compliance. Methods: A deep learning model was trained for DME detection using the EyePacs dataset. Data was randomly assigned, by participant, into development (n= 14,246) and validation (n= 1,583) sets. Analysis was conducted at the single image, eye, and patient levels. Model performance was evaluated using sensitivity, specificity, and AUC. Findings: At the patient level, sensitivity of 0.901 (CI 95% 0.879-0.917), specificity of 0.900 (CI 95% 0.883-0.911), and AUC of 0.962 (CI 95% 0.955-0.968) were achieved. At the image level, sensitivity of 0.889 (CI 95% 0.878-0.900), specificity of 0.889 (CI 95% 0.877-0.900), and AUC of 0.954 (CI 95% 0.949-0.959) were achieved. At the eye level, sensitivity of 0.905 (CI 95% 0.890- 0.920), specificity of 0.902 (CI 95% 0.890-0.913), and AUC of 0.964 (CI 95% 0.958-0.969) were achieved. Interpretation: DME can be detected from color fundus imaging with high performance on all analysis metrics. Automatic DME detection may simplify screening, leading to more comprehensive screening for diabetic patients. Further prospective studies are necessary.

No abstract

BackgroundDiabetic retinopathy (DR) is a leading cause of blindness in adults worldwide. Artificial intelligence (AI) with autonomous deep learning algorithms has been increasingly used in retinal image analysis, particularly for the screening of referrable DR. An established treatment for proliferative DR is panretinal or focal laser photocoagulation. Training autonomous models to discern laser patterns can be important in disease management and follow-up.MethodsA deep learning model was trained for laser treatment detection using the EyePACs dataset. Data was randomly assigned, by participant, into development (n=18 945) and validation (n=2105) sets. Analysis was conducted at the single image, eye, and patient levels. The model was then used to filter input for three independent AI models for retinal indications; changes in model efficacy were measured using area under the receiver operating characteristic curve (AUC) and mean absolute error (MAE).ResultsOn the task of laser photocoagulation detection: AUCs of 0.981, 0.95, and 0.979 were achieved at the patient, image, and eye levels, respectively. When analysing independent models, efficacy was shown to improve across the board after filtering. Diabetic macular oedema detection on images with artefacts was AUC 0.932 vs AUC 0.955 on those without. Participant sex detection on images with artefacts was AUC 0.872 vs AUC 0.922 on those without. Participant age detection on images with artefacts was MAE 5.33 vs MAE 3.81 on those without.ConclusionThe proposed model for laser treatment detection achieved high performance on all analysis metrics and has been demonstrated to positively affect the efficacy of different AI models, suggesting that laser detection can generally improve AI-powered applications for fundus images.

BackgroundDiabetic retinopathy is a leading cause of blindness in adults worldwide. AI with autonomous deep learning algorithms has been increasingly used in the analysis of retinal images particularly for the screening of referrable DR. An established treatment for proliferative DR is pan-retinal or focal laser photocoagulation. Training AI autonomous models to discern laser patterns can be important in disease management and follow-up.MethodsA deep learning model was trained for laser treatment detection using the EyePACs dataset. Data was randomly assigned, by participant, into development (n= 18,945) and validation (n= 2,105) sets. Analysis was conducted at the single image, eye, and patient levels. The model was then used to filter input images for three independent AI models for various retinal indications, and changes in model efficacy were measured using AUC and MAE.FindingsOn the task of laser photocoagulation detection: AUC of 0.981 (CI 95% 0.971-0.87) was achieved at the patient level. AUC of 0.950 (CI 95% 0.943-0.956) was achieved at the image level. AUC of 0.979 (CI 95% 0.972-0.984) was achieved at the eye level.When analyzing independent AI models, efficacy was shown to improve across the board on images of untreated eyes. DME detection on images with artifacts was AUC 0.932 (CI 95% 0.905-0.951) vs. AUC 0.955 (CI 95% 0.948-0.961) on those without. Participant sex detection on images with artifacts was AUC 0.872 (CI 95% 0.830-0.903) compared to AUC 0.922 (CI 95% 0.916-0.927) on those without. Participant age detection on images with artifacts was MAE 5.33 vs. MAE 3.81 on those without.InterpretationThe proposed model for laser treatment detection achieved high performance on all analysis metrics and has been demonstrated to positively affect the efficacy of different AI models, suggesting that laser detection can generally improve AI powered applications for fundus images.FundingProvided by AEYE Health Inc.