AMD-UPerNet: a tool for retinal layer and fluid assessment in age-related macular degeneration

This study aims to develop a rapid and precise OCT segmentation model that simultaneously delineates retinal layers and fluid regions in age-related macular degeneration (AMD). AMD-UPerNet was proposed as the segmentation model designed to simultaneously delineate retinal layers and fluid compartments in OCT images. Key innovations include: (1) Swin Transformer was adopted as the backbone to extract multi-scale features and model global contextual relationships; (2) Content-Aware ReAssembly of FEatures (CARAFE) module was introduced into feature pyramid network to improve up-sampling and mitigate feature loss; (3) The combination of cross-entropy and Dice loss was used to handle class imbalance and reduce mis-segmentation of fluid regions. Compared to the models such as PSPNet, DANet, OCNet, DenseASPP, SAM, MedSAM, and SegFormer, AMD-UPerNet achieved superior performance, with a Pixel Accuracy (PA) of 98.51%, Mean Pixel Accuracy (MPA) of 87.49%, Mean Precision (MPre) of 89.40%, and Mean Intersection over Union (MIoU) of 80.53%. Unlike previous methods that segment retinal layers and fluid regions separately, our model uses a joint framework that enables simultaneous optimization and contextual interaction. By using the anatomical continuity of retinal layers as structural priors, AMD-UPerNet improves fluid localization and enhances robustness against ambiguous boundaries and low-contrast regions. These results highlight AMD-UPerNet’s potential to improve OCT segmentation accuracy and efficiency, facilitating early diagnosis and optimized treatment for AMD.