RetinalFRNet: retinal vessel segmentation in OCTA images using the feature reconstruction network.

OBJECTIVES

Automated segmentation of retinal blood vessels in optical coherence tomography angiography (OCTA) images is essential for early diagnosis of ocular diseases such as diabetic retinopathy, myopia, and macular degeneration. This study evaluates the performance of the Retinal Feature Reconstruction Network (RetinalFRNet), a full-resolution deep learning architecture designed for OCTA retinal vessel segmentation.

METHODS

RetinalFRNet integrates recurrent neural network modules within a ConvNeXt backbone while preserving full spatial resolution without downsampling. Its performance was compared with five algorithms - fuzzy C-means (FCM), U-Net, ResUnet, Vision Transformer (ViT), and VM-Unet - using three publicly available datasets (OCTA-3 mm, OCTA-6 mm, and ROSSA). All models were trained using the Adam optimizer for 100 epochs and evaluated using seven standard metrics including Accuracy, Dice coefficient, Sensitivity, and mean Intersection over Union (IoU).

RESULTS

RetinalFRNet achieved the highest overall performance across all datasets, reaching a Dice coefficient of 91.72 % and mean IoU of 84.93 % on the ROSSA dataset, improving performance by up to 30.66 % over the FCM baseline.

CONCLUSIONS

RetinalFRNet demonstrates superior accuracy and robustness for OCTA retinal vessel segmentation. Its downsampling-free architecture enables improved detection of fine vascular structures critical for ophthalmic diagnosis. Further multi-center validation is recommended prior to clinical deployment.