Dual-stream frequency-spatial fusion and feature calibration for robust thyroid nodule segmentation

Accurate segmentation of thyroid nodules in ultrasound (US) imaging is a critical prerequisite for automated clinical diagnosis. However, this task remains highly challenging due to inherent speckle noise, low tissue contrast, and severely ambiguous lesion boundaries. Conventional Convolutional Neural Networks (CNNs) primarily focus on local spatial domains, which inevitably leads to the progressive loss of high-frequency boundary cues and the introduction of semantic noise via standard skip connections. To address these bottlenecks, we propose a novel architecture named Dual-Stream Frequency-Spatial Fusion Network (DS-FSFNet). Specifically, we introduce a Frequency-Spatial Fusion (FSF) module at the network bottleneck. By leveraging the Fast Fourier Transform (FFT), the FSF module explicitly extracts and integrates high-frequency structural representations, effectively compensating for the spatial blurring effect of deep convolutions. Furthermore, a Spatial-Channel Feature Calibration (SFC) module is embedded within the skip connections to adaptively filter out semantic noise using dual-attention mechanisms prior to feature decoding. Extensive experiments on three challenging thyroid US datasets (DDTI, TG3K, and TN3K) demonstrate that DS-FSFNet achieves competitive segmentation performance compared with representative medical image segmentation methods. In addition, evaluation on the CVC-ClinicDB dataset, which involves a substantially different medical imaging task, provides further evidence of the transferability of the proposed architecture across different segmentation scenarios.