GCCA-trans: a hybrid segmentation network with guided criss-cross attention for early-stage head and neck tumors in CT images

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurate lesion segmentation in medical images is essential for disease assessment and treatment planning. Most head and neck tumors can be clinically cured when treated at an early stage, making early detection highly beneficial for guiding effective therapy. However, research on early-stage head and neck tumor segmentation remains limited due to the scarcity of dedicated datasets. To address this gap, we collect and annotate an early-stage head and neck tumor dataset (ES-HNT), comprising contrast-enhanced CT scans from 138 patients. To address the challenges of segmenting small, low-contrast tumors, we propose a CNN–Transformer hybrid architecture, GCCA-Trans, built around two components: a Guided Cyclic Cross-Attention (GCCA) module embedded at the skip connections, which captures global contextual relationships to strengthen the representation of small tumor regions, and a Squeeze-Enhanced Axial Attention (SEA) module, which replaces standard multi-head self-attention in the Transformer encoder to preserve fine-grained local detail alongside global context modeling. Experimental results on the ES-HNT and SegRap2023 datasets show that GCCA-Trans achieves consistent improvements over existing state-of-the-art methods under the same experimental settings.