Comparison of the Segment Anything Model and U-Net-Based Architectures for Retinal Fluid Segmentation Using the AROI Dataset

Retinal fluid segmentation from optical coherence tomography (OCT) images is important for diagnosing and monitoring retinal diseases such as neovascular age-related macular degeneration. U-Net-based architectures are the standard for this task, providing reliable performance on domain-specific datasets. Recently, foundation models such as the Segment Anything Model (SAM) have emerged as general-purpose segmentation tools that operate in a prompt-based, zero-shot manner, but their performance on challenging medical imaging tasks remains underexplored. This study presents a comparative evaluation of SAM and U-Net-based architectures for retinal fluid segmentation on the AROI dataset, a publicly available benchmark containing multi-class annotations of retinal layers and pathological fluids. SAM is applied in a zero-shot setting using automatically generated box prompts derived from annotated fluid regions. Its performance is compared against U-Net and Attention U-Net models using the Dice coefficient and Intersection over Union (IoU). SAM achieved the best results for both intraretinal fluid (IRF) and subretinal fluid (SRF), while U-Net-based models were more consistent for pigment epithelial detachment (PED), benefiting from anatomical context learned during training. We analyze the applicability, strengths, and limitations of a general-purpose segmentation model in a clinically challenging OCT scenario and highlight differences relative to specialized medical architectures.