WaveGate-nnU-Net: Frequency-Domain Inclusion Preserves Segmentation Accuracy under Reduced Training Data in Nasal and Paranasal Sinus CT
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background and Objectives
Automatic 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.
Methods
We 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.
Results
Under 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 = 1.9×10⁻⁶) and was corroborated by average surface distance.
Conclusions
The practical value of this combined approach lies primarily in within-distribution data efficiency, rather than in maximizing peak segmentation accuracy.
