Wavelet-Driven Cross-Domain Consistency for Mixed-Supervised 3D Tumor Segmentation

Accurate 3D tumor segmentation is vital for clinical decision-making, but pixel-wise annotation is expensive. Mixed-supervised learn-ing combines a small set of mask labels with abundant box labels, yetreconstruction-kernel heterogeneity creates large domain gaps. We pro-pose WCC4MS, a jointly optimized lesion-centered framework that mod-els kernel-related frequency shifts via a 3D discrete wavelet transform.The input volume is decomposed into low- and high-frequency com-ponents, and the scaled high-frequency coefficients generate syntheticsmooth and sharp variants. A dual-branch architecture learns from maskand box annotations with task-specific decoders, while a cross-branchKL-divergence constraint enforces semantic consistency. An auxiliarydomain-aware head encourages quality-sensitive features and improvesrobustness to domain shift. At inference, only the segmentation branch isused on candidate-centered patches; the box branch is used only duringtraining to exploit weak annotations. Experiments across three mask-annotated CT datasets and one box-only dataset show consistent gainsin segmentation accuracy and cross-domain generalization compared toprior mixed- and weakly supervised methods.