FreqFuseNet: Scale-Normalized Dual-Frequency Fusion for Thin-Wall Head-and-Neck OAR Segmentation
- 1Chang Gung University and Linkou Chang Gung Memorial Hospital, Taoyuan, Taiwan
Abstract
Accurate segmentation of thin-wall organs-at-risk (OARs) in head-and-neck computed tomography (CT) is challenging because small contour deviations can substantially affect boundary agreement. We present FreqFuseNet, a segmentation architecture with a fast Fourier transform (FFT) primary branch and a frequency channel attention network (FcaNet) residual branch that explicitly aligns heterogeneous frequency-feature scales before bounded residual fusion. An initialization-time analysis revealed an approximately 863-fold activation-scale disparity between the two branches, corresponding to an approximately 43-fold unintended residual-to-primary magnitude ratio under naive fusion with a nominal coefficient of 0.05. FreqFuseNet uses per-sample, per-channel scale alignment to preserve the intended relative amplitude without additional normalization parameters. We evaluated 10 thin-wall OARs on SegRap2023 (84 training, 18 validation, and 18 testing CT cases) under a standardized binary per-OAR region-of-interest (ROI) protocol. In two training seeds, FreqFuseNet achieved Dice of 0.849 and 0.843, 95th-percentile Hausdorff distance (HD95) of 0.824 and 0.823 mm, and surface Dice at 1 mm of 0.959 and 0.957. In case-level paired comparisons for the primary seed (Wilcoxon signed-rank test, not corrected for multiple comparisons), FreqFuseNet showed significant improvements over 3D U-Net and MedNeXt-S, whereas differences versus an alternative learned fusion design (FixedFusion) and SegResNet were not statistically significant. These findings identify a practical feature-scale calibration issue in dual-frequency segmentation and demonstrate competitive boundary-sensitive performance of the proposed architecture within the evaluated ROI-based setting.
Related
- Boundary-aware multi-scale residual learning with hybrid radiomic–deep fusion for lung tumor segmentation and classification
- UAV image segmentation in complex mining areas by integrating spatially adaptive feature modulation and multi-scale context fusion modules
- MTF-Net: a brain tumor segmentation network with multi-scale transformer and feature fusion enhancement
- AMF-U-Net: an adaptive multimodal fusion residual attention 3D U-Net for boundary-aware brain tumour segmentation
- Structure-preserving lightweight skin lesion segmentation with dual-resolution attention and boundary-aware fusion
BibTeX
@misc{chen2026freqfusenet,
title = {{FreqFuseNet: Scale-Normalized Dual-Frequency Fusion for Thin-Wall Head-and-Neck OAR Segmentation}},
author = {W.-Y. Chen and G.-Y. Lin and S.-Y. Wan},
howpublished = {medRxiv},
year = {2026},
month = jul,
doi = {10.64898/2026.07.09.26357642},
url = {https://www.medrxiv.org/content/10.64898/2026.07.09.26357642v3},
note = {CC BY},
}
RIS
TY - UNPB
TI - FreqFuseNet: Scale-Normalized Dual-Frequency Fusion for Thin-Wall Head-and-Neck OAR Segmentation
AU - W.-Y. Chen
AU - G.-Y. Lin
AU - S.-Y. Wan
PY - 2026
DA - 2026/07/09
PB - medRxiv
DO - 10.64898/2026.07.09.26357642
UR - https://www.medrxiv.org/content/10.64898/2026.07.09.26357642v3
AB - Accurate segmentation of thin-wall organs-at-risk (OARs) in head-and-neck computed tomography (CT) is challenging because small contour deviations can substantially affect boundary agreement. We present FreqFuseNet, a segmentation architecture with a fast Fourier transform (FFT) primary branch and a frequency channel attention network (FcaNet) residual branch that explicitly aligns heterogeneous frequency-feature scales before bounded residual fusion. An initialization-time analysis revealed an approximately 863-fold activation-scale disparity between the two branches, corresponding to an approximately 43-fold unintended residual-to-primary magnitude ratio under naive fusion with a nominal coefficient of 0.05. FreqFuseNet uses per-sample, per-channel scale alignment to preserve the intended relative amplitude without additional normalization parameters. We evaluated 10 thin-wall OARs on SegRap2023 (84 training, 18 validation, and 18 testing CT cases) under a standardized binary per-OAR region-of-interest (ROI) protocol. In two training seeds, FreqFuseNet achieved Dice of 0.849 and 0.843, 95th-percentile Hausdorff distance (HD95) of 0.824 and 0.823 mm, and surface Dice at 1 mm of 0.959 and 0.957. In case-level paired comparisons for the primary seed (Wilcoxon signed-rank test, not corrected for multiple comparisons), FreqFuseNet showed significant improvements over 3D U-Net and MedNeXt-S, whereas differences versus an alternative learned fusion design (FixedFusion) and SegResNet were not statistically significant. These findings identify a practical feature-scale calibration issue in dual-frequency segmentation and demonstrate competitive boundary-sensitive performance of the proposed architecture within the evaluated ROI-based setting.
ER -