Lightweight Frequency-Gated Enhancement of nnU-Net Improves Data Efficiency in Nasal Sinus CT Segmentation
- 1Chang Gung University
Abstract
Background and ObjectivesAutomatic segmentation of the nasal cavity and paranasal sinuses from CT aids diagnosis and surgical planning, but clinical datasets in this domain remain small, which can affect training stability and evaluation validity. This study investigates whether frequency-domain representation learning can be incorporated, in a lightweight form, into the self-configuring nnU-Net framework, and whether the resulting performance improvement survives rigorous statistical validation.
MethodsWe constructed an nnU-Net with frequency inclusion, adding a multi-scale wavelet branch and a cross-domain attention module at the network bottleneck following a previously reported frequency-spatial dual-stream formulation, and corrected a zero-initialization gating deadlock in that formulation, using the corrected model (v2) as a common base. On this base, we evaluated three variants: X1, an input-conditioned learnable spectral gate applied to 24 wavelet sub-bands; X2, which adds a boundary-distance training objective; and WaveGate-nnU-Net (hereafter WaveGate), the proposed model combining both. Experiments used the 130-volume NasalSeg CT dataset, under a full 91/19/20 protocol and a 3-fold cross-validation in which the training set for each fold was reduced by approximately 19%, from 91 to 73-74 cases.
ResultsUnder the full protocol, all variants showed a small but statistically detectable improvement over a matched nnU-Net baseline. WaveGate achieved a Dice score of 95.78% versus 95.59% for the baseline, an improvement of 0.19 percentage points (95% CI [0.09, 0.29]; Holm-adjusted p = 0.004), and WaveGate also reduced average surface distance (ASD) from 0.252 mm to 0.235 mm. WaveGate was statistically equivalent to either individual component within a margin of {+/-}0.10 percentage points, suggesting a shared performance ceiling rather than a complementary gain. Under reduced training data, baseline Dice dropped by 2.71 percentage points (95.59% to 92.88%), while WaveGate remained essentially unchanged (95.78% to 95.75%), a cross-fold advantage of +2.88 percentage points (fold-level 95% CI [0.80, 4.96]) that held across all 20 test cases (sign test, p = 1.9x10-) and was corroborated by average surface distance.
ConclusionsThe practical value of this combined approach lies primarily in within-distribution data efficiency, rather than in maximizing peak segmentation accuracy.
Related
- Structure-preserving lightweight skin lesion segmentation with dual-resolution attention and boundary-aware fusion
- Efficient and real-time polyp segmentation: An empirical evaluation of lightweight Mamba integration strategies
- A “stepped” flexible vacuum-assisted ureteral access sheath improves stone clearance efficiency: experimental results from an ex vivo porcine model
- MTF-Net: a brain tumor segmentation network with multi-scale transformer and feature fusion enhancement
- UniMedSeg: Unified Multi-Modal Medical Image Segmentation via Uncertainty Guided Encoding and Frequency-Spatial Dual-Prompt Decoding
BibTeX
@misc{chen2026lightweight,
title = {{Lightweight Frequency-Gated Enhancement of nnU-Net Improves Data Efficiency in Nasal Sinus CT Segmentation}},
author = {Y.-C. Chen and S.-Y. Wan},
howpublished = {medRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.10.26362809},
url = {https://www.medrxiv.org/content/10.64898/2026.09.10.26362809v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Lightweight Frequency-Gated Enhancement of nnU-Net Improves Data Efficiency in Nasal Sinus CT Segmentation
AU - Y.-C. Chen
AU - S.-Y. Wan
PY - 2026
DA - 2026/09/10
PB - medRxiv
DO - 10.64898/2026.09.10.26362809
UR - https://www.medrxiv.org/content/10.64898/2026.09.10.26362809v1
AB - Background and ObjectivesAutomatic segmentation of the nasal cavity and paranasal sinuses from CT aids diagnosis and surgical planning, but clinical datasets in this domain remain small, which can affect training stability and evaluation validity. This study investigates whether frequency-domain representation learning can be incorporated, in a lightweight form, into the self-configuring nnU-Net framework, and whether the resulting performance improvement survives rigorous statistical validation. MethodsWe constructed an nnU-Net with frequency inclusion, adding a multi-scale wavelet branch and a cross-domain attention module at the network bottleneck following a previously reported frequency-spatial dual-stream formulation, and corrected a zero-initialization gating deadlock in that formulation, using the corrected model (v2) as a common base. On this base, we evaluated three variants: X1, an input-conditioned learnable spectral gate applied to 24 wavelet sub-bands; X2, which adds a boundary-distance training objective; and WaveGate-nnU-Net (hereafter WaveGate), the proposed model combining both. Experiments used the 130-volume NasalSeg CT dataset, under a full 91/19/20 protocol and a 3-fold cross-validation in which the training set for each fold was reduced by approximately 19%, from 91 to 73-74 cases. ResultsUnder the full protocol, all variants showed a small but statistically detectable improvement over a matched nnU-Net baseline. WaveGate achieved a Dice score of 95.78% versus 95.59% for the baseline, an improvement of 0.19 percentage points (95% CI [0.09, 0.29]; Holm-adjusted p = 0.004), and WaveGate also reduced average surface distance (ASD) from 0.252 mm to 0.235 mm. WaveGate was statistically equivalent to either individual component within a margin of {+/-}0.10 percentage points, suggesting a shared performance ceiling rather than a complementary gain. Under reduced training data, baseline Dice dropped by 2.71 percentage points (95.59% to 92.88%), while WaveGate remained essentially unchanged (95.78% to 95.75%), a cross-fold advantage of +2.88 percentage points (fold-level 95% CI [0.80, 4.96]) that held across all 20 test cases (sign test, p = 1.9x10-) and was corroborated by average surface distance. ConclusionsThe practical value of this combined approach lies primarily in within-distribution data efficiency, rather than in maximizing peak segmentation accuracy.
ER -