Low-Power Retinal Vessel Segmentation Using Frequency-Domain Transformers

Retinal vessel segmentation plays a crucial role in early diagnosis of diabetic retinopathy, glaucoma, hypertension, and other ocular diseases. However, existing deep learning models often require high computational resources, making them unsuitable for real-time deployment on portable or IoT-based ophthalmic devices. This paper proposes a low-power retinal vessel segmentation framework using a Frequency-Domain Transformer (FDT) that efficiently captures elongated vascular structures with minimal computational overhead. The proposed method transforms retinal images into the frequency domain using a lightweight Discrete Fourier Transform module, enabling global context extraction with reduced spatial computations. A compressed multi-head spectral attention mechanism further enhances thin and low-contrast vessel detection while maintaining low memory footprint. Experimental results on DRIVE, STARE, and CHASE_DB1 datasets demonstrate that the proposed FDT model achieves competitive accuracy while consuming significantly less power (up to 40% reduction) compared to conventional CNN-based models such as U-Net and ResNet variants. The model’s efficiency and accuracy make it suitable for handheld fundus cameras, tele-ophthalmology systems, and on-device screening applications.