FMUNet: Frequency Modulation-Based UNet for Octa Segmentation

In analyzing optical coherence tomography angiography (OCTA) images, segmentation is a crucial step in the development of computer-aided diagnosis systems for detecting retinal diseases. However, existing approaches often suffer from limitations, such as a fixed receptive field or the inability to preserve the salient structure of the vascular system, which hampers the accuracy of disease detection. To address these challenges, we introduce the Fourier Transform. Building on this concept, we propose FMUNet: Frequency Modulation-Based U-Net for OCTA Segmentation. This model is designed to achieve an image-sized receptive field while capturing global dependencies, thereby enhancing segmentation performance. Specifically, we apply the Fast Fourier Transform (FFT) to feature maps across multiple scales, followed by a lightweight block that learns attention masks from the frequency representations obtained post-FFT. This approach emphasizes transferable frequency components that are beneficial for representing the vascular system while suppressing less informative components. Experimental results demonstrate that FMUNet achieves superior performance compared to state-of-the-art models, with mean Dice and mean IoU scores of 91.39% and 84.15%, respectively, highlighting its robustness and enhanced capability in OCTA segmentation.