FA‐SSNet: A Frequency‐Aware Semi‐Supervised Network for Thyroid Nodule Segmentation in Ultrasound Images
IET Image Processing · 2026-01 · 20 (1)
DOI: 10.1049/ipr2.70518 ↗Kurz gefasst
Fachartikel in der Zeitschrift IET Image Processing, online erschienen am 16.11.2026. Verfasst von Rowzatul Zannath Prerona und 1 weiteren Autor:innen. Die Arbeit nennt 49 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Accurate segmentation of thyroid nodules in ultrasound images remains a clinically demanding task because of edge blurring caused by speckle noise, wide variability in nodule size and the scarcity of expert‐annotated training data. Existing deep‐learning approaches operate entirely in the spatial domain and rely on fully supervised training, leaving the physical root cause of boundary degradation—the attenuation of high‐frequency signal components—unaddressed. We propose FA‐SSNet, a frequency‐aware semi‐supervised network that tackles these limitations through three main contributions. (i) A dual‐branch encoder that processes each ultrasound image simultaneously through a Swin Transformer spatial branch and a frequency branch built on 2D discrete wavelet transform and fast Fourier transform, enabling explicit capture of boundary‐defining high‐frequency information. (ii) A novel cross‐domain attention fusion (CDAF) module that lets spatial semantic features and frequency edge features guide each other through bidirectional gating and cross‐attention, producing a richer joint representation. (iii) A frequency‐weighted composite loss (FWCL)—a spatial‐domain composite loss combining Dice, binary cross‐entropy and deep‐supervision objective whose design is directly motivated by the frequency‐domain features extracted by the frequency branch, coupled with a mean‐teacher semi‐supervised strategy that exploits unlabelled scans to reduce annotation cost. An explicit FFT‐domain boundary term was evaluated during development but excluded from the final loss owing to numerical instability under mixed‐precision training. Evaluated on a combined pool of 7101 images assembled from three public benchmarks (TN3K, DDTI, TG3K; a further 614 official TN3K test images were held out and not used in this study), FA‐SSNet achieves a Dice similarity coefficient of 0.8485 (IoU = 0.7699, AUC‐ROC = 0.9911, ECE = 0.0137), representing a 6.73 percentage‐point improvement over the strongest reproduced baseline, Swin U‐Net (DSC = 0.7812), evaluated under the same metric definitions on a differently‐sourced test set. Ablation studies confirm complementary gains from each component (0.08, 0.07 and 0.05 DSC, respectively). The Wilcoxon signed‐rank test () and a 95% bootstrap confidence interval of establish statistical significance across all datasets.
Zitieren
APA
Prerona, R. Z., & Rahman, M. A. (2026). FA‐SSNet: A Frequency‐Aware Semi‐Supervised Network for Thyroid Nodule Segmentation in Ultrasound Images. IET Image Processing, 20(1). https://doi.org/10.1049/ipr2.70518
Vancouver
Prerona RZ, Rahman MA. FA‐SSNet: A Frequency‐Aware Semi‐Supervised Network for Thyroid Nodule Segmentation in Ultrasound Images. IET Image Processing. 2026;20(1). doi:10.1049/ipr2.70518
BibTeX
@article{Prerona2026,
author = {Prerona, Rowzatul Zannath and Rahman, Mohammad Amanour},
title = {FA‐SSNet: A Frequency‐Aware Semi‐Supervised Network for Thyroid Nodule Segmentation in Ultrasound Images},
journal = {IET Image Processing},
year = {2026},
volume = {20},
number = {1},
doi = {10.1049/ipr2.70518}
}
EUVIMED