SEEM-OCTA: Geometry-Informed Interactive Refinement for Parameter-Efficient OCTA Vessel Segmentation

Accurate segmentation of retinal vasculature from Optical Coherence Tomography Angiography (OCTA) is critical for reliable assessment of diabetic retinopathy and macular degeneration. Although foundation segmentation models have shown strong performance on natural images, their application to OCTA imaging remains limited due to domain mismatch, high adaptation cost, and ineffective interaction strategies for thin vascular structures. To address these challenges, we propose SEEM-OCTA, a parameter-efficient and topology-aware framework for OCTA vessel segmentation. Specifically, we employ Low-Rank Adaptation (LoRA) to enable efficient fine-tuning of a large foundation model with minimal trainable parameters. Furthermore, we utilize Geometry-Informed Interactive Refinement (GIIR), which guides the selection of interaction points based on vessel topology to enhance segmentation accuracy with fewer user inputs. We evaluate the proposed approach on the OCTA-500 dataset and demonstrate that SEEM-OCTA achieves a Dice score of 0.911 using only 3.4 ± 2.3 clicks, corresponding to a 39\% reduction in interaction effort compared to the baseline. Overall, our results demonstrate the effectiveness of SEEM-OCTA as a practical and accurate solution for OCTA-based vessel segmentation. Our code is publicly available at https://github.com/jaafaralghabban/seem-octa.