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

Conference: ECCV 2026
Paper: ECCV Official
Full Cache: ../paper_cache/ECCV2026/eccv-3143.txt
Area: Medical Imaging
Keywords: 3D tumor segmentation, mixed-supervised learning, wavelet transform, cross-domain consistency, domain generalization

TL;DR¶

To tackle CT reconstruction-kernel heterogeneity and multi-source domain shifts, this paper proposes WCC4MS, a wavelet-driven cross-domain consistency framework for mixed-supervised 3D tumor segmentation that synthesizes smooth/sharp variants via 3D wavelet modulation and regularizes a dual-branch network with cross-task KL consistency and domain-aware objectives.

Background & Motivation¶

Accurate 3D CT tumor segmentation plays an indispensable role in quantifying lesion morphology, monitoring disease progression, and personalizing oncological treatment plans. While fully supervised deep learning models trained on high-quality pixel-wise masks have achieved remarkable success, obtaining dense voxel annotations for 3D CT volumes is notoriously labor-intensive and requires rigorous cross-validation by experienced radiologists. To mitigate the annotation burden, weakly supervised segmentation relying solely on cost-effective bounding boxes has drawn significant attention. However, purely weakly supervised models often struggle to resolve ambiguous and irregular lesion boundaries, failing to satisfy the rigorous accuracy demands of clinical translation. Mixed-supervised learning has therefore emerged as an attractive paradigm, aiming to leverage a small fraction of dense mask annotations alongside abundant bounding-box annotations to train a unified segmentation model.

Nonetheless, existing mixed-supervision approaches face severe domain shift and imaging heterogeneity across datasets. In clinical practice and multi-center benchmarks, datasets with dense masks and weak bounding boxes frequently stem from diverse scanner vendors and clinical cohorts with distinct CT reconstruction kernels. High-noise sharp kernels preserve crisp tissue boundaries, whereas low-noise smooth kernels soften fine-grained textures. Due to patient privacy concerns, public CT datasets routinely strip raw DICOM metadata, rendering physical inverse filtering via the Fourier transform infeasible. Furthermore, standard image quality assessment tasks focus on detecting global distribution discrepancies, whereas tumor segmentation demands sharp, localized boundary classification. Conventional domain adaptation or semi-supervised frameworks struggle to reconcile these divergent objectives.

To resolve this fundamental tension, the authors approach the challenge through multiscale frequency-domain modulation. While raw reconstruction metadata is missing, the physical effect of different kernels fundamentally manifests as the amplification or suppression of high-frequency components. By decomposing 3D volumes into subbands using a 3D discrete wavelet transform and scaling the high-frequency detail coefficients, one can synthesize smooth and sharp variants without corrupting the underlying anatomical morphology. Core idea: scale high-frequency detail coefficients via 3D discrete wavelet transforms to synthesize smooth and sharp domain variants, train a shared-backbone dual-branch architecture on mask and box annotations, and regularize feature representations via cross-branch KL consistency and an auxiliary domain classification objective, thereby achieving highly robust cross-domain 3D tumor segmentation under reconstruction-kernel heterogeneity.

Method¶

Overall Architecture¶

WCC4MS follows an end-to-end, lesion-centered dual-branch framework designed to operate on candidate-centered patches (\(96 \times 96 \times 96\)), matching standard clinical 3D computer-aided detection (CAD) workflows where suspicious regions are localized first and then refined. The input volume is first processed by the KernelWave module, which performs multi-level wavelet decomposition (MWD) to decouple low-frequency anatomical structures from high-frequency directional details, subsequently scaling high-frequency coefficients to generate synthetic smooth and sharp volume pairs.

These perturbed volume variants are fed into a shared 3D feature encoder coupled with task-specific decoding branches: a segmentation branch supervised by dense ground-truth masks, and an auxiliary detection branch supervised by bounding-box annotations. To reinforce robustness against high-frequency perturbations, a symmetric KL-divergence consistency constraint is enforced across the dual-branch output probability distributions. Furthermore, a lightweight domain classification head is attached to the shared encoder to explicitly classify data source provenance as an auxiliary regularizer, encouraging the encoder to retain quality-sensitive cues rather than forcibly discarding domain specifics. At test time, only the segmentation branch is retained for inference, requiring zero additional computational overhead.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    In["Input 3D Lesion Patch X<br/>(96×96×96)"] --> KW["KernelWave Frequency Modulation<br/>MWD + HF scaling into smooth/sharp variants"]
    KW --> SharedEnc["Shared 3D Backbone<br/>(3D UNet / 3D UNetr)"]
    SharedEnc --> DMS["Dual-Branch Mixed Supervision<br/>Segmentation Branch (Dice) + Auxiliary Det Branch (Det)"]
    DMS --> KL["Cross-Branch Semantic Consistency<br/>KL divergence across output distributions"]
    SharedEnc --> DC["Auxiliary Domain-Aware Regularizer<br/>Predicts dataset source, preserves quality cues"]
    DMS --> Out["Inference Output<br/>Fine 3D tumor segmentation mask only"]

Key Designs¶

1. KernelWave Frequency Modulation: Metadata-Free Reconstruction Kernel Simulation

CT reconstruction kernels act as frequency-domain filters, selectively boosting or suppressing high-frequency bands. Because patient privacy strips raw reconstruction parameters, exact inverse Fourier compensation cannot be performed. KernelWave resolves this by applying a 3D multi-level wavelet decomposition (MWD). At decomposition level \(v\), an input volume \(X \in \mathbb{R}^{D \times H \times W}\) is decomposed into one low-frequency approximation subband \(\mathcal{X}_{LLL}^{(v)}\) and seven directional high-frequency detail subbands \(\mathcal{X}_{\epsilon}^{(v)}\) where \(\epsilon \in \{LLH, LHL, HLL, LHH, HLH, HHL, HHH\}\). The low-frequency subband captures stable anatomical structure, while high-frequency subbands encode noise and boundary sharpness.

To simulate kernel-induced frequency shifts while preserving anatomical geometry, KernelWave keeps the low-frequency subband untouched and modulates high-frequency detail coefficients via a scaling factor \(\delta > 1\). Specifically, sharp variants amplify high frequencies as \(\widetilde{\mathcal{X}}_{\epsilon}^{(v)} = \delta \cdot \mathcal{X}_{\epsilon}^{(v)}\), whereas smooth counterparts attenuate them as \(\widetilde{\mathcal{X}}_{\epsilon}^{(v)} = \frac{1}{\delta} \cdot \mathcal{X}_{\epsilon}^{(v)}\). The modified subbands are then reconstructed via multi-level wavelet reconstruction (MWR):

\[\widetilde{X} = \mathrm{MWR}\left(\mathcal{X}_{LLL}^{(V)}, \{\widetilde{\mathcal{X}}_{\epsilon}^{(v)}\}\right)\]

This mechanism faithfully mirrors the spectral energy shifts between standard smooth kernels (e.g., B31f) and sharp kernels (e.g., B60f) without relying on heuristic spatial blur or handcrafted noise.

2. Dual-Branch Mixed Supervision: Decoupling Mask and Bounding Box Information

Directly concatenating weak bounding-box objectives with dense pixel-level mask supervision within a single decoder often degrades fine-grained boundary sensitivity due to label granularity conflicts. The DMS module decouples the learning objectives into task-dedicated branches over a shared 3D encoder. The segmentation branch is supervised with soft Dice loss on full-resolution mask labels:

\[\mathcal{L}_{\mathrm{dsc}} = 1 - \frac{2 \sum_{i} p_i g_i + \sigma}{\sum_{i} p_i^2 + \sum_{i} g_i^2 + \sigma}\]

Simultaneously, the auxiliary detection branch explicitly capitalizes on the deterministic spatial constraints of bounding boxes (center position and 3D dimensions) via a multi-task detection objective \(\mathcal{L}_{\mathrm{det}} = \mathcal{L}_{\mathrm{cls}} + \mathcal{L}_{\mathrm{reg}}\), combining classification cross-entropy with 3D bounding-box IoU regression. This auxiliary supervision forces the shared representation to maintain strong spatial localization awareness even when pixel masks are extremely scarce.

3. Cross-Branch Semantic Consistency: Distribution Alignment Under Frequency Shift

When exposed to KernelWave-synthesized smooth and sharp pairs, the model must maintain structural semantic invariance despite substantial high-frequency appearance divergence. For the dual predictions \(p^{(1)}\) and \(p^{(2)}\) generated across the branches, a voxel-wise symmetric KL-divergence loss is enforced across the 3D volume domain \(\Omega\):

\[\mathcal{L}_{\mathrm{kl}} = \frac{1}{|\Omega|} \sum_{i \in \Omega} \left[ p^{(1)}_i \log\left(\frac{p^{(1)}_i + \sigma}{p^{(2)}_i + \sigma}\right) + (1 - p^{(1)}_i) \log\left(\frac{1 - p^{(1)}_i + \sigma}{1 - p^{(2)}_i + \sigma}\right) \right]\]

This objective penalizes discordant confidence assignments between smooth and sharp inputs, stabilizing gradient trajectories and preventing the segmentation branch from overfitting to spurious high-frequency artifacts.

4. Auxiliary Domain-Aware Regularizer: Non-Adversarial Representation Learning

Conventional domain generalization frequently relies on adversarial training to strip domain-specific cues and learn "domain-invariant" features. In medical CT imaging, however, domain-specific properties like reconstruction sharpness and scanner noise are inextricably linked with lesion texture and edge clarity; eliminating them indiscriminately harms segmentation precision. The authors introduce a lightweight domain classification head on top of the shared encoder, supervised by cross-entropy across known training data sources \(C\):

\[\mathcal{L}_{\mathrm{dc}} = - \sum_{c=1}^{C} y_{k}^{(c)} \log \hat{y}_{k}^{(c)}\]

Acting as an explicit regularizer rather than an adversarial discriminator, this objective guides the encoder to retain image-quality and acquisition-specific cues, empowering the downstream decoder to adapt to varying kernel regimes rather than blindly discarding critical high-frequency boundary evidence.

Loss & Training¶

The entire WCC4MS framework is trained end-to-end using a joint composite objective:

\[\mathcal{L}_{\mathrm{total}} = \lambda_1 \mathcal{L}_{\mathrm{dsc}} + \lambda_2 \mathcal{L}_{\mathrm{det}} + \lambda_3 \mathcal{L}_{\mathrm{kl}} + \lambda_4 \mathcal{L}_{\mathrm{dc}}\]

Except for the cross-branch KL divergence term, all losses represent the sum of contributions from both branches. The optimal hyperparameter weighting is established as \(\lambda_1 = 1, \lambda_2 = 1, \lambda_3 = 0.5, \lambda_4 = 0.1\). The model is optimized using SGD (initial learning rate 0.01, momentum 0.9, weight decay 1e-4) paired with a cosine annealing scheduler, trained with batch size 8 for 80 epochs on two NVIDIA RTX 4090 GPUs.

Key Experimental Results¶

Main Results¶

Experiments are evaluated across three mask-annotated thoracic CT benchmarks (LIDC-IDRI, MSD-Lung, LNDb) and one large-scale box-only auxiliary dataset (PN9). The mask-to-box supervision ratio is set to 1:2. Performance is quantified using the Dice Similarity Coefficient (DSC, %) and 95% Hausdorff Distance (HD95, mm). Comparative results using a 3D UNet backbone are summarized below:

Type	Methods	LIDC-IDRI (DSC↑ / HD95↓)	MSD-Lung (DSC↑ / HD95↓)	LNDb (DSC↑ / HD95↓)
Fully Supervised	3D UNet (Full Labels)	77.26 / 1.92	56.46 / 7.31	63.24 / 8.59
Weakly Supervised	SANPNet	65.31 / 2.86	46.83 / 14.46	57.76 / 14.20
Weakly Supervised	MonoBox	68.03 / 2.75	47.21 / 15.56	60.25 / 11.46
Weakly Supervised	ReliableMD	68.54 / 2.73	46.14 / 14.19	61.21 / 12.32
Weakly Supervised	BoxInst	66.86 / 2.78	48.95 / 14.93	60.47 / 11.85
Mixed Supervised	MIST	77.45 / 1.90	56.18 / 8.14	64.00 / 8.27
Mixed Supervised	ConfKD	73.85 / 2.47	52.02 / 12.26	63.18 / 9.87
Mixed Supervised	MixSegNet	78.19 / 1.75	55.36 / 8.13	63.72 / 8.39
Mixed Supervised	WSL4MIS	77.62 / 1.73	57.01 / 7.49	63.99 / 8.63
Mixed Supervised	MSDN	72.44 / 2.53	51.94 / 12.13	62.25 / 9.76
Ours	WCC4MS	80.32 / 1.52	57.93 / 7.06	65.84 / 7.98

When equipped with a 3D UNetr backbone, WCC4MS reaches 81.16% (HD95 1.43), 60.02% (HD95 8.21), and 67.90% (HD95 7.51) across the three datasets, consistently outperforming prior SOTA methods. In leave-one-domain-out domain generalization, WCC4MS attains an average DSC of 59.88%, clearly outperforming GTVSeg (57.93%) and the unadapted DeepAll baseline (51.66%). In cross-modality validation across four brain MRI benchmarks, WCC4MS also achieves superior segmentation performance on BraTS (85.69% vs 83.71%), MSD-BT (83.04% vs 81.08%), and FeTS (83.62% vs 82.41%).

Ablation Study¶

Ablation experiments on LIDC-IDRI isolating the four core components—KernelWave (KW), KL consistency (KL), domain classification regularizer (DC), and auxiliary detection branch (AD)—are presented below:

No.	KW (Wavelet)	KL (Consistency)	DC (Domain Reg)	AD (Aux Det)	DSC (%) ↑	HD95 (mm) ↓	p-value	Note
1					63.17	3.74	0.009	Baseline segmentation branch only
2				✔	72.95	2.39	0.015	Adding auxiliary detection branch
3	✔	✔	✔		66.27	3.10	0.001	Omitting bounding-box supervision
4	✔	✔		✔	76.89	2.03	0.005	Omitting domain classification head
5	✔		✔	✔	78.05	1.95	0.036	Omitting cross-branch KL consistency
6		✔	✔	✔	75.46	2.17	0.018	Omitting wavelet frequency modulation
7 (Full)	✔	✔	✔	✔	80.32	1.52	-	Full WCC4MS framework

Key Findings¶

Crucial Role and Saturation of Weak Box Supervision: Disabling the auxiliary detection branch drops DSC severely from 80.32% to 66.27% (-14.05%), confirming that bounding boxes provide indispensable spatial priors. Analyzing the mask-to-box ratio reveals that performance peaks at 1:2; pushing the ratio past 1:4 to 1:5 causes a sharp drop, as excessive bounding-box gradients divert attention away from fine boundary textures.
Physical Fidelity of KernelWave: Disabling KernelWave reduces DSC by 4.86% (80.32% vs 75.46%). Spectral validation on real multi-kernel CT scans (RKPN with B31f/B60f pairs) demonstrates a spectral correlation of \(r = 0.71\) between KernelWave shifts and real kernel variations, with high-frequency energy ratios of 1.50 vs 1.44, substantially outperforming Gaussian blur (79.48%) and Laplacian filtering (76.98%).
Resilience to Box Label Noise: When random 20% spatial perturbation is injected into bounding-box centers and dimensions, the top competing baseline drops steeply to 72.9% DSC, whereas WCC4MS maintains 77.6% DSC, highlighting robust tolerance against imprecise annotations.

Highlights & Insights¶

Decoupling Anatomy and Noise via Wavelet Subbands: Utilizing 3D discrete wavelets to preserve low-frequency anatomical geometry while modulating high-frequency details elegantly circumvents the loss of raw DICOM reconstruction metadata in public datasets.
Domain Retention Over Domain Erasure: Instead of blindly forcing domain-invariant representations via adversarial min-max games, supervising a domain classification head preserves image-quality cues, allowing decoders to condition on specific quality regimes.
Zero-Cost Inference Deployment: Auxiliary detection heads and wavelet synthesis modules are active strictly during training; the deployment phase requires only the segmentation branch on localized 3D patches, making it fully plug-and-play for real-world CAD pipelines.

Limitations & Future Work¶

Dependency on Upstream Candidate Localization: The model presumes an initial CAD detection stage has extracted 3D lesion patches (\(96 \times 96 \times 96\)); severe localization offsets or false-negative omissions upstream cannot be recovered by the segmentation head alone.
Heuristic Wavelet Gain Parameter: The scaling factor \(\delta = 4\) is determined via empirical search rather than dynamically conditioned on localized CT tissue contrast.
Future Directions: The authors intend to validate the frequency consistency framework across modalities such as ultrasound and PET, and explore extensions to ultra-sparse annotations including scribbles and extreme points.

vs MIST / WSL4MIS (Mixed-Supervised Medical Segmentation): Prior mixed-supervised works rely heavily on heuristic pseudo-labeling or self-training over shared backbones, leaving them susceptible to error propagation under domain shifts. WCC4MS explicitly models kernel-induced frequency shifts and stabilizes learning through decoupled detection and KL consistency.
vs GTVSeg / MKCNet (Medical Domain Generalization): Existing domain generalization frameworks typically align feature distributions globally, which often discards informative high-frequency edge textures. WCC4MS leverages domain-aware regularizers alongside wavelet modulation to preserve boundary sharpness across multi-center cohorts.

Rating¶

Novelty: ⭐⭐⭐⭐☆ [Creative use of 3D wavelets to simulate CT reconstruction kernels and non-adversarial domain quality regularization.]
Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Extensive evaluations across 3 CT benchmarks, 1 large box dataset, real multi-kernel validation, and cross-modality brain MRI.]
Writing Quality: ⭐⭐⭐⭐⭐ [Clear structural organization, thorough mathematical formulations, and sharp alignment with clinical challenges.]
Value: ⭐⭐⭐⭐⭐ [Provides a practical, highly robust blueprint for clinical CAD segmentation leveraging abundant low-cost bounding boxes.]