Classification of range of OCT-angiography capillary density using multichannel deep learning models in diabetic retinopathy, aging macular degeneration, and radiation retinopathy

OCT-angiography (OCTA) is an increasingly important image modality for the diagnosis and management of various retinal vascular diseases. However, manual analysis is time-consuming and subjective. Automated segmentation using deep learning can provide consistent and objective measurements, improving capillary density assessments. Studies have shown that automated quantification of capillary density is beneficial, but these studies have primarily classified areas of retina as having normal or abnormal capillary density, and are unable to classify various degrees of capillary density. In this study, an automated method for segmenting microvascular density regions into four categories (normal, mild, moderate, and severe) was developed using a nnU-Net framework. We proposed four models with different input combinations: Model 1 (OCTA only), Model 2 (OCTA + foveal avascular zone [FAZ]), Model 3 (OCTA + large vessel tree [LVT]), and Model 4 (OCTA + FAZ + LVT) to determine whether increasing the number of inputs improved prediction accuracy. The dataset comprised of 50 training and 47 test images, labeled by two independent experts. Dataset 1 (train: 50, test: 15) was annotated by Expert 1, while Dataset 2 (test: 32) was annotated by both experts. The predicted labels were compared to the manual reference labels using Dice coefficients. Although no significant differences between Expert 1 and the models were identified, visual inspection revealed that Model 4 (3-channel input) occasionally produced more consistent and uniform results. ANOVA tests compared the Dice coefficients for Expert 1, Expert 2, and Model 4 and found significant differences only in the normal category (p-value: 0.036), but Tukey’s HSD test found no significant differences between each pair comparison. This study demonstrates that a three-channel input model (OCTA + FAZ + LVT) provides more objective and consistent segmentation of capillary density in OCTA images. The model’s performance, comparable to human experts, represents a significant advancement in automated OCTA image analysis, offering an alternative to manual assessments.