Dnq-unet: a two-level fusion framework for few-shot domain adaptation in cervical cancer CTV segmentation

To develop and validate a two-level hierarchical fusion architecture enabling efficient few-shot domain adaptation for cervical cancer clinical target volume (CTV) auto-segmentation across different institutional imaging and contouring settings. We propose DNQ-UNet, a dual-encoder architecture with two fusion levels: spatially adaptive normalization (SPADE) for shallow appearance alignment and cross-attention for deep semantic transfer. A progressive two-stage training strategy with grouped learning rates was employed. Few-shot adaptation used 3, 5, 10, 20, 40, and 60 fine-tuning cases, and additional 3D fine-tuning baselines, including nnU-Net V2, SwinUNETR, and MedNeXt, were evaluated under identical target-domain test settings. Ablation studies assessed each fusion component. On the target-domain test set, zero-shot DSC was [Formula: see text]. With 3, 5, and 10 fine-tuning cases, DSC improved to [Formula: see text], [Formula: see text], and [Formula: see text], respectively. DNQ-UNet achieved the highest DSC among the compared 3D baselines at 5-shot (0.830 vs. 0.805-0.817) and 10-shot (0.838 vs. 0.823-0.833) settings. Ablation showed significant DSC reductions after removing either fusion level or using a single U-Net (all [Formula: see text]). The proposed framework enables efficient few-shot domain adaptation for cervical cancer CTV segmentation using 5-10 annotated cases, providing a locally adapted contour-initialization approach that may reduce annotation and repeated correction burden during cross-institutional deployment.