FreqFuseNet: Resolving Feature-Scale Mismatch in Dual-Frequency Fusion for Thin-Wall Head-and-Neck OAR Segmentation

Accurate segmentation of thin-wall organs-at-risk (OARs)-the cochlea, vestibular semicircular canals, internal auditory canal, tympanic cavity, and middle ear-is clinically relevant for head-and-neck radiotherapy planning, yet these small, thin-wall structures remain among the most challenging targets for automated delineation. Dual-frequency feature fusion is a promising direction for boundary-sensitive representation, but under the investigated FP16 FFT-FcaNet setting, we observe an approximately 863-fold activation-scale mismatch between the FFT and FcaNet branches, causing a nominal 5 percent residual coefficient to behave as an approximately 43-fold dominant term. We propose FreqFuseNet, which resolves this mismatch by normalizing the FcaNet branch to the FFT activation scale before residual injection with a fixed low-amplitude coefficient (beta = 0.05), restoring beta as an interpretable 5 percent residual-amplitude coefficient relative to the FFT feature scale. Under a controlled binary per-OAR ROI protocol on the SegRap2023 head-and-neck CT benchmark across 10 clinically prioritized thin-wall OARs, FreqFuseNet achieves Dice of 0.849, HD95 of 0.824 mm, and SDice@1mm of 0.959 in the primary seed, with comparable performance in an independent second seed (Dice 0.843, HD95 0.823 mm). FreqFuseNet yields statistically significant case-level aggregate improvements over 3D U-Net and MedNeXt-S (Wilcoxon p < 0.01 and p < 0.05, respectively), using only 29.7 million parameters versus 414.6 million for the full wavelet baseline.