DAA-UNet: A Dense Connectivity and Atrous Spatial Pyramid Pooling Attention UNet Model for Retinal Optical Coherence Tomography Fluid Segmentation

Abstract

Retinal optical coherence tomography (OCT) fluid segmentation is a vital tool for diagnosing and treating various ophthalmic diseases. Based on clinical manifestations, retinal fluid accumulation is classified into three categories: intraretinal fluid (IRF), subretinal fluid (SRF), and pigment epithelial detachment (PED). PED is primarily associated with diabetic macular edema (DME). In contrast, IRF and SRF play critical roles in diagnosing age-related macular degeneration (AMD) and retinal vein occlusion (RVO). To address challenges posed by variations in OCT imaging devices, as well as the varying sizes, irregular shapes, and blurred boundaries of fluid accumulation areas, this study proposes DAA-UNet, an enhanced UNet architecture. The proposed model incorporates dense connectivity, Atrous Spatial Pyramid Pooling (ASPP), and attention gate (AG) in the paths of UNet. Dense connectivity expands the model’s depth, whereas ASPP facilitates the extraction of multiscale image features. The AG emphasize critical spatial location information, improving the model’s ability to distinguish different fluid accumulation types. Experimental results on the MICCAI 2017 RETOUCH challenge dataset showed that DAA-UNet demonstrates superior performance, with a mean Dice Similarity Coefficient (mDSC) of 90.2%, 91.6%, and 90.5% on cirrus, spectralis, and topcon devices, respectively. These results outperform existing models, including UNet, SFU, Attention-UNet, Deeplabv3+, nnUNet RASPP, and MsTGANet.