Abstract
Purpose :
We previously developed a deep-learning system (DLS) to identify diabetic macular ischemia (DMI) on superficial capillary plexus(SCP) and deep capillary plexus(DCP) of optical coherence tomography angiography (OCT-A) images. We aim to further investigate whether the DLS-based DMI classification provides prognostic values on diabetic retinopathy (DR) progression and diabetic macular oedema (DME) development in a cohort of patients with diabetes mellitus (DM).
Methods :
This is a retrospective longitudinal study with 293 eyes from 164 patients with DM being followed up for at least 4 years.Image quality assessment and DMI assessment of all OCT-A images were first performed by the previously developed DL system. The presence of DMI was defined as images exhibiting disruption of fovea avascular zone (FAZ) or/and additional areas of capillary non-perfusion in the macula accorded with ETDRS protocols. Cox proportional-hazards model was used to evaluate the relationship of the binary DMI classification (presence or absence of DMI) to DR progression and DME development.
Results :
Over a median follow-up of 54.09 ±4.67 months, 89 eyes (33.97%) had DR progression, and 34 eyes (12.98%) developed DME. SCP-DMI classification outcome (hazard ratio [HR], 3.451; 95% confidence interval [CI], 2.062 to 5.776; p <0.001) and DCP-DMI classification outcome (HR, 4.992; 95% CI, 2.884 to 8.645; p <0.001) were significantly associated with DR progression; while only DCP-DMI classification outcome (HR, 1.758; 95% CI 0.753 to 4.122; p <0.001) was associated with DME development after adjusting for age, duration of diabetes, fasting glucose, glycated hemoglobin, mean arterial blood pressure, baseline DR severity, average GCIPL thickness, and smoking status at baseline. Compared with the model based on identified risk factors alone, the addition of DMI classification outcomes improved the predictive discrimination of DR progression (SCP-DMI, C-statistics 0.741 vs. 0.696, p < 0.001; DCP-DMI, C-statistics 0.772 vs. 0.696, p < 0.01) and DME development (DCP-DMI, C-statistics 0.702 vs. 0.663, p = 0.076) in diabetic eyes.
Conclusions :
The DL-based DMI outcomes demonstrated prognostic values for DR progression and DME development.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.