FPE-Unet: Thyroid Nodule Segmentation via Frequency and Position-Aware

Accurate segmentation of thyroid nodules in ultrasound images is essential for diagnosis but remains challenging due to noise, size variability, and spatial irregularity. We propose FPE-Unet, which includes: (1) Frequency Domain Processing (FDP) to reduce noise and enhance boundaries; (2) Position Attention with Receptive Field (PARF) for multi-scale, spatiallyaware detection; (3) Efficient Bottleneck (E-Bottleneck) combining depthwise and asymmetric convolutions to strengthen feature representation. On TN3K and DDTI datasets, FPE-Unet achieves Dice scores of 85.17% and 82.35%, outperforming 14 existing methods. Ablation studies confirm the value of each component. FPE-Unet demonstrates strong robustness and generalization for clinical use.