Precise anatomical delineation in ultrasound imaging using echo-spectral interactive networks

Accurate ultrasound (US) segmentation is challenging because speckle noise, low contrast, and ambiguous tissue interfaces often obscure anatomical boundaries. Many conventional CNN- and Transformer-based models operate primarily in the spatial domain and may not explicitly incorporate boundary-related spectral cues.

To develop a spatial-frequency interactive segmentation framework designed to support boundary delineation in US images.

We introduce the Echo-Spectral Interactive Transformer (ESI-Former) block, which uses a Spectrum-Directed Self-Attention (SDSA) module to inject differentiable DCT-derived spectral representations into self-attention. We further introduce a Skip-Aware Attention Fusion Module (SAFM) to adaptively fuse encoder skip features with decoder representations. Based on these components, we construct a U-shaped segmentation architecture termed SF-EchoNet.

SF-EchoNet was evaluated on the TN3K thyroid nodule segmentation dataset and the benign subset of the Breast Ultrasound Images (BUSI) dataset. It achieved run-level mean mIoU scores of 88.01% and 79.58%, with numerically lower run-level mean HD95 values of 41.31 and 47.22, respectively, than the compared methods. Efficiency evaluation showed 12.24M parameters, 9.98 ms latency, and approximately 100 FPS on an NVIDIA RTX A6000 GPU.

On the two evaluated public datasets, SF-EchoNet achieved higher run-level mean overlap scores and numerically lower run-level mean HD95 values than the compared methods. Paired analyses supported the overlap differences on both datasets, whereas the magnitude of the mean HD95 difference on the BUSI benign subset remained uncertain. Generalizability to other datasets, acquisition settings, and downstream clinical tasks requires further validation.