Abstract
Purpose :
To validate the performance of an artificial-intelligence(AI) based software for the detection and quantification of major OCT biomarkers in eyes affected by diabetic macular edema(DME).
Methods :
After training a deep learning algorithm using normal and diabetic OCT eyes images, a specific tool was developed for the automatic detection and quantification of OCT biomarkers: intraretinal and subretinal fluids volume(IRF and SRF), hyperreflective retinal foci(HRF), integrity of external limiting membrane(ELM) and ellipsoid zone(ELZ). The distribution of IRF was also automatically quantified. Full map scan was analyzed by the AI automatic software and compared to clinical evaluation. The agreement between the software and clinical evaluation was calculated for: HRF, SRF and both ELM and EZ, and the accuracy of volumes quantification for IRF and SRF was evaluated.
Results :
Three hundred and three eyes were enrolled. Mean central subfield thickness was 386.5±130.2micron. The mean volume of IRF detected by the software was 0.898±1.367 mm3, with a mean of:13.9±18.0%, 34.4±21.9%, 51.3±30.2% of IRF in the 1, 3, 6 ETDRS mm, respectively, and mean IRF density (%/relative surface area) of 0.088±0.114, 0.047±0.068 and 0.025±0.043 in the 1, 3, 6 ETDRS mm, respectively. SRF (relevant when ≥0.002 mm3) was detected in 43 eyes by the software, with a mean volume of 0.111±0.191mm3. The agreement for SRF detection was almost perfect (k=0.831,AC1=0.948). The automatic quantification was defined clinically accurate in 289(95.38%) eyes for IRF and 287(94.72) for SRF. The mean number of HRF detected by the software was 71.9±22.8 vs 71.9±22.7 detected by clinical evaluation. The ICC between the clinical evaluation and the software was excellent(0.97). In the Bland-Altmann plot almost all measured differences were into the range (2SD), with a mean difference between the clinical and automatic count of 0.03±5.277. Kappa Inter-rater agreement (95% confidence interval) was 0.934 for ELM integrity, and 0.936 for EZ integrity.
Conclusions :
Accurate, repeatable detection and quantification of OCT biomarkers in DME are currently mandatory to diagnose, prognosticate and follow treatment response. Our fully validated AI based tool allows the clinicians to identify and quantify these clinical parameters offering an objective way of precisely diagnosing and following, in a fully quantitative way, DME eyes.
This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.