Dense-PMSFNet: DenseNet Pyramidal Multi-Scale Fusion Network for Retinal Vasculature and FAZ Segmentation in OCTA Images

Abstract

Diabetic retinopathy (DR) is a condition that leads to damage to the retina due to high blood sugar levels in the eyes caused by prolonged diabetes. People of working age in developing countries are particularly at risk of developing diabetic retinopathy, which causes permanent vision loss in diabetic patients. Diagnosis involves maintaining the current level of vision of the patient, as the disease can cause blindness. Optical Coherence Tomography Angiography (OCTA) is an advanced eye imaging technique that offers a detailed view of the structures of retinal blood vessels and effectively treats various eye conditions. Therefore, it is crucial to accurately identify and separate the capillary, artery, vein, and Foveal Avascular Zone (FAZ) from OCTA images. The DenseNet Pyramidal MultiScale Fusion Network (Dense-PMSFNet) is introduced in this article as a new segmentation network based on deep learning. It utilizes the DenseNet Encoder to integrate local semantic contextual information, the Multi-Scale Pyramidal Fusion Module (MSPFM) to effectively merge local features at different scales, and Deep Fusion to enhance the representation of multiscale decoder outputs. Through extensive experimentation on four distinct segmentation tasks involving capillary, artery, vein, and FAZ from OCTA images, it has demonstrated superior performance for the quantitative metrics such as Dice coefficient (DSC), Accuracy (ACC), Intersection Over Union (IoU), Specificity (SP) and Sensitivity (SE) in comparison to state-of-the-art methods.