Pattern Analysis and Applications Компьютерное зрение и распознавание образов

GBKA-Net: a gated Kolmogorov–Arnold network with boundary-guided multi-scale attention for thyroid nodule segmentation

Ali, Hassan, Chen, Zhe
Springer Nature · 2026-10-31 · DOI 10.1007/s10044-026-01797-7

Аннотация

Thyroid nodules must be delineated precisely in ultrasound scans before computer-aided systems can reliably support diagnosis, but several properties of the modality make this difficult: speckle corrupts the signal, contrast between nodule and surrounding tissue is often weak, nodules vary widely in texture and shape, and their margins are frequently ill-defined. Motivated by these difficulties, we propose GBKA-Net (Gated Boundary-guided Kolmogorov–Arnold Attention Network), a segmentation architecture that extends the Kolmogorov–Arnold network paradigm with two complementary gating mechanisms acting at the bottleneck and along the decoder skip pathways. At the bottleneck, the Gated Kolmogorov–Arnold Multi-Scale Attention (GKMSA) module couples spline-based KAN feature modeling with grouped multi-scale attention, using a global context map from the coarse-scale branch to gate the fine-scale branch so that global anatomical structure suppresses speckle and local artifacts before attention weights are computed. Along the decoder, the Gated Boundary-Aware Fusion (GBAF) module replaces plain additive skip connections by combining encoder features, fine boundary cues, and reverse-attention evidence from adjacent higher levels into a single spatial gate that retains encoder detail only where boundary or region evidence is present. A dedicated boundary branch produces explicit edge representations, and the network is optimized with a deep-supervision objective combining boundary-weighted cross-entropy and IoU terms over the segmentation heads together with a boundary loss on the edge head. Experiments on five public datasets demonstrate the effectiveness of GBKA-Net. On the thyroid ultrasound datasets TN3K and DDTI, it achieves Dice scores of 92.14% and 94.87% and IoU scores of 85.48% and 90.29%, respectively. Cross-domain evaluation on BUSI and UDIAT attains Dice scores of 81.0% and 82.0%, and zero-shot evaluation on the independent Stanford Thyroid Ultrasound Cine-Clip cohort preserves the ranking over U-KAN and U-Net, indicating cross-organ transfer within B-mode ultrasound and robustness under domain shift.

Читать у издателя →