FRNet V2: a lightweight full-resolution convolutional neural network for OCTA vessel segmentation

Optical coherence tomography angiography (OCTA) is an advanced non-invasive imaging technique that can generate three-dimensional images of retinal and choroidal vessels. It is of great value in the diagnosis and monitoring of a variety of ophthalmic diseases. However, most existing methods for blood vessel segmentation in OCTA images rely on an encoder–decoder architecture. This architecture typically involves a large number of parameters and leads to slower inference speeds. To address these challenges and improve segmentation efficiency, this paper proposes a lightweight full-resolution convolutional neural network named FRNet V2 for blood vessel segmentation in OCTA images. FRNet V2 combines the ConvNeXt V2 architecture with deep separable convolution and introduces a recursive mechanism. This mechanism enhances feature representation while reducing the amount of model parameters and computational complexity. In addition, we design a lightweight hybrid adaptive attention mechanism (DWAM) that further improves the segmentation accuracy of the model through the combination of channel self-attention blocks and spatial self-attention blocks. The experimental results show that on two well-known retinal image datasets (OCTA-500 and ROSSA), FRNet V2 can achieve Dice coefficients and accuracy comparable to other methods while reducing the number of parameters by more than 90%. In conclusion, FRNet V2 provides an efficient and lightweight solution for fast and accurate OCTA image blood vessel segmentation in resource-constrained environments, offering strong support for clinical applications.