acceptodds
Under review as a conference paper at ICLR 2027

WDFNet: Adaptive Frequency Guidance for Diffusion Based Ultrasound Image Classification

Abstract

Existing label diffusion classifiers rely primarily on deep semantic image conditioning, in which frequency information is learned only implicitly. Although deep networks can capture frequency dependent patterns, prior studies have reported frequency biases in learned representations and shown that explicit frequency modeling can complement conventional spatial features in visual and medical image analysis. This issue is particularly relevant to ultrasound, where high frequency responses may represent fine anatomical boundaries and tissue texture, but may also reflect speckle related patterns and acquisition variation. We propose WDFNet, an adaptive frequency conditioning framework for label diffusion classification that explicitly introduces wavelet reconstructed high frequency evidence into the image guidance pathway. Its Adaptive Frequency Guide (AFG) selects complementary frequency responses according to semantic context, while its Frequency Alignment Module (FAM) adjusts the fused condition according to the current noisy label, diffusion timestep, and auxiliary class prior. Experiments on six medical image datasets show consistent improvements over the evaluated discriminative and diffusion based classifiers. Repeated runs, capacity matched controls, fixed fusion controls, paired label noise analyses, and controlled corruption experiments further support the contribution of adaptive frequency conditioning rather than label diffusion or additional branch capacity alone.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical image analysis

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vKneOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors investigate medical image classification, particularly ultrasound images, where fine anatomical boundaries and texture can be informative but speckle and acquisition artifacts can also influence prediction.

In their method and architecture desig, they have used the semantic features and frequency features together to address this problem. A ConvNeXt encoder extracts image features and produces an auxiliary class prediction. The authors add a wavelet branch that extracts high-frequency image details and uses a learned gate to combine those details with semantic features.

They introduce two architectural components, Adaptive Frequency Guide (AFG) and Frequency Alignment Module (FAM).

The authors evaluate their architecture on six medical imaging datasets and report improvements over discriminative classifiers and DiffMIC baselines. Their central argument is that explicitly representing frequency information, selecting it adaptively, and using it during label denoising improves classification accuracy.

Critical strengths

The motivation and architecture they have designed is aligned with ultrasound imaging: The authors recognize that high-frequency information can contain both useful anatomical detail and nuisance variation such as speckle. Their approach avoids assuming that all high-frequency information should be preserved or suppressed.

Experiments: The beat the other baseless and capacity-matched comparison against a second branch receiving the original image is a well-designed comparison. This helps address the concern that performance improves simply because the model has more parameters. The authors test ultrasound, CT, dermoscopy, and MRI datasets. It demonstrates applicability beyond one ultrasound benchmark.

Critical weaknesses

Figure 1 is a little bit confusing. Your method also use semantic features while I can interpret that from the box b that you use only frequency features. (The captions is correct and mention it but the schematics a little bit misleading.)

The necessity of label diffusion is not clear: The authors demonstrate that frequency features improve their diffusion classifier. However, this does not establish that diffusion is necessary to exploit those features.

Consider two models:

\text{Model A:}\quad& (F_s,F_h)\rightarrow\mathrm{AFG} \rightarrow\mathrm{MLP}\rightarrow\hat y,\\ \text{Model B:}\quad& (F_s,F_h)\rightarrow\mathrm{AFG} \rightarrow\mathrm{FAM}\rightarrow \text{Label diffusion}. \end{aligned}$$ If Model A obtains comparable results, the contribution may be a frequency-enhanced classifier rather than a meaningful advancement in diffusion modeling. Limited methodological novelty: Prior work has already combined wavelet representations and adaptive attention for medical classification and ultrasound segmentation, including FEANet and Wavelet Attention Fusion. Combining different parts together to get the best results is a valid contribution but every parts need an ablation why this specific features and architecture needed. A convincing experiment should compare WDFNet with and without FAM, semantic-only guidance with FAM, and frequency-only guidance with FAM.
Questions for authors

Is label diffusion necessary for the proposed improvements? Could the authors compare WDFNet against a conventional discriminative classifier using the same semantic encoder, wavelet features, and AFG module? This would help determine whether the improvements come from frequency-aware representation learning or specifically from label diffusion.

What is the independent contribution of FAM? Since FAM is retained in all configurations in Table 3, could the authors provide an ablation removing FAM while keeping the other components unchanged? How much does FAM actually improve classification performance?

What is the computational benefit of using label diffusion for this classification task? Given that WDFNet uses 100 reverse sampling steps, could the authors report inference latency, FLOPs, and performance under fewer sampling steps? How does its accuracy–efficiency tradeoff compare with frequency-enhanced discriminative classifiers?

3 comments

Thank you for recognizing the ultrasound motivation, capacity-matched spatial control, and evaluation across imaging modalities. We clarify the semantic/frequency streams and report prediction-pathway, FAM and single-stream controls together with inference costs. WDFNet gains 7.71 Accuracy points over AFG+MLP under speckle, and 10-step DDIM takes 1.09 ms/image with matching predictions across sampling budgets.

1. Semantic and frequency streams

Thank you for pointing out the schematic ambiguity. Both streams enter the condition through

Ffuse=Fs+G(Fs,Fh)⊙Fh.F_{\mathrm{fuse}}=F_s+G(F_s,F_h)\odot F_h.

The semantic feature FsF_s supplies the residual and helps compute the gate that selects frequency details.

We will revise Figure 1 to show the semantic residual and gated frequency branch explicitly.

2. Added value of label diffusion

Following your suggestion, we train the same ConvNeXt–Haar–AFG image architecture with an MLP and end-to-end cross entropy.

Prediction pathwayClean Acc / macro F1 (%)Speckle Acc / macro F1 (%)
AFG + end-to-end MLP93.14 / 92.5385.49 / 84.52
WDFNet96.00 / 95.5093.20 / 91.83

These comparisons use the paper's training settings. WDFNet's Accuracy/F1 gains are 2.86/2.97 points on clean images and 7.71/7.31 points under speckle. The larger speckle margin highlights the added prediction pathway's value when image conditions are perturbed.

3. Independent FAM ablation

Thank you for suggesting this control. We replace FAM with fixed addition, ht=Vg+zth_t=V_g+z_t, retaining AFG and the noisy-label/timestep/prior inputs.

BUSI conditioningAccuracy (%)Macro F1 (%)
Fixed addition (w/o FAM)89.90 ± 1.0788.62 ± 1.66
FAM92.16 ± 1.8090.23 ± 1.25

Values are mean ± SD over five training seeds, using the paper's training settings and original BUSI split. FAM raises mean Accuracy/F1 by 2.26/1.61 points.

4. Inference cost and sampling steps

Thank you for asking about practical cost. Measurements cover the full model on RTX 5090/FP32, batch size 16. Latency is the median of ten runs after three warmups, excluding I/O, preprocessing, and transfer.

Model / modeParameters (M)Counted GFLOPs / imagems / image
USFM85.7735.140.99
AFG + end-to-end MLP221.2811.040.49
WDFNet, original 100-index sampler, uncached239.74181.4117.73
WDFNet, original 100-index sampler, cached239.7440.959.63

GFLOPs are profiler-counted lower bounds. Caching reuses the image condition across steps, with zero output difference in the parity check. This reduces counted GFLOPs by 77.4% and measured latency by 45.7% for the original sampler.

With DDIM indices spanning the full diffusion schedule, 10, 20, 50 and 100 steps give identical class predictions:

DDIM stepsCounted GFLOPs / imagems / image
1013.761.09
2016.781.70
5025.853.55
10040.966.65

Thus, 10 steps retain the measured classification performance at lower inference cost.

5. Frequency conditioning and branch ablations

Our focus is adaptive frequency conditioning during label denoising. In Table 3 of the paper, replacing the raw-image branch with wavelet details improves Accuracy/F1 by 2.1/2.4 points, with the same branch and gating architectures and FAM retained.

Following your suggestion, we also compare single-stream image conditions:

Image conditionNew PMG2000 clean Accuracy (%)Macro F1 (%)
Semantic-only94.7294.21
Frequency-only94.9994.32
Combined96.0095.50

These new experiments use the paper's training settings. All retain FAM and the auxiliary prior; stream labels refer to the image condition. Combining streams improves Accuracy by 1.28 points over semantic-only and 1.01 over frequency-only, supporting complementary image cues.

We will add the prediction-pathway and image-stream controls and a compact accuracy–cost comparison to the main text, with detailed sampling-budget measurements in the appendix.

Thanks for your proper response and the additional experiments for answering my concerns. FAM's importance addressed well. I still have concern regarding the label diffusion as they are performing same over 10 to 100 steps so essentially the Diffusion model learned an easy distribution and when map can not reach the same performance is interesting. Maybe it is too late to ask to check the training process and make sure fairness there. Also I can not see the edits in the uploaded paper yet. I need to read the paper one final time for my final decision. please update the final versions with all of the necessary edits.

Thank you for your thoughtful follow-up and for recognizing the contribution of FAM.

Following your suggestion, we audited the AFG+MLP implementation, training logs, checkpoints, and gradient propagation. We verified that its ConvNeXt–Haar–AFG branch produces identical features to WDFNet when using the same weights and inputs, and that all active trainable parameter tensors received optimizer updates. AFG+MLP completed 600 training epochs, with its cross-entropy loss decreasing from 1.2983 to 0.00217 and approaching zero toward the end of training. These observations indicate that the baseline was adequately fitted to its training objective.

Both models share the data partitions, preprocessing, augmentation, pretrained encoder initialization, and training budget. Their auxiliary optimization protocols differ, reflecting their respective training objectives.

We also independently repeated DDIM inference at all four sampling budgets. Although the predicted classes remain identical, the continuous label vectors differ across all test images between 10 and 100 steps, with a maximum absolute difference of 0.05218. Thus, the observed consistency reflects stable classification decisions rather than identical denoising outputs. Moreover, a short reverse trajectory does not imply equivalence between conditional label-noise estimation and direct cross-entropy training.

Following CARD and DiffMIC, our contribution focuses on frequency conditioning within label diffusion. The AFG+MLP comparison supports the empirical benefit of the complete prediction pathway, rather than isolating the effect of diffusion alone.

We have revised the manuscript to incorporate the additional experiments. Thank you again for your constructive suggestions and careful evaluation.

Reviewer_M6n9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

WDFNet builds on the DiffMIC-style label diffusion framework (CARD), which generates class-label vectors using image features and an auxiliary prior. It introduces two key changes to the image-conditioning process. First, the Adaptive Frequency Guidance (AFG) module uses a Haar wavelet transform to extract image details and combines them with ConvNeXt features through a learnable gate. Second, the Feature Alignment Module (FAM) adjusts the image features based on the noisy label, timestep, and auxiliary prior. The method is evaluated on six datasets, including three ultrasound datasets (PMG2000, UCPM2200, and BUSI), where it outperforms several existing CNN-, transformer-, and diffusion-based methods. Additional experiments on PMG2000 investigate the contributions of each module, the choice of wavelet settings, robustness to image noise, and label-noise prediction using a smaller model.

Critical strengths

I think that the main strength of this paper is its clarity and presentation. It is easy to follow and reads well, with the main contributions clearly explained and the experiments presented in a straightforward way. The method is evaluated on six datasets, providing a broad range of results across different medical imaging tasks. The ablation studies are also interesting, particularly those exploring different wavelet transforms, levels, and basis functions, as they help clarify the role of frequency information and provide useful insights into the model's performance.

Critical weaknesses

My main concern is the strength of the experimental evidence. Most baseline results are taken from previous papers instead of being reproduced under the same settings, and results on datasets other than PMG2000 are based on a single run. Some improvements are also very small, making it difficult to determine whether they are statistically meaningful. In addition, the method is not compared with more recent state-of-the-art representation learning approaches for ultrasound, such as USFM, which makes it harder to assess its performance against current methods.

Another concern is the novelty of the proposed modules, particularly FAM. AFG mainly adds high-frequency image details through a gating mechanism, so its contribution may be somewhat incremental. Moreover, FAM is not evaluated separately, making it difficult to determine how much it contributes to the overall performance.

Finally, an important comparison is missing. The paper does not evaluate a simpler ConvNeXt model with AFG but without label diffusion. This would help clarify whether the diffusion framework itself provides additional benefits or whether the improvements mainly come from the frequency-based features.

Questions for authors

I generally like the paper, particularly its clarity, presentation, and experimental design. However, I have a few concerns that make it difficult for me to recommend acceptance at this stage. I would appreciate clarification on the following points:

  • Can the authors provide evidence that the reported improvements are statistically meaningful?
  • Why is WDFNet not compared with recent ultrasound representation learning approaches, such as USFM? How does it perform against these methods?
  • Can the authors provide an ablation study that isolates the contribution of FAM and demonstrates whether it provides a meaningful improvement?
  • Can the authors compare WDFNet with a ConvNeXt model using AFG but without label diffusion to clarify the added value of the diffusion framework?

I find the paper interesting and would be willing to increase my assessment if these concerns are adequately addressed. However, without further evidence on these points, I find it difficult to recommend acceptance at this stage.

2 comments

Thank you for recognizing the clear presentation, six-dataset evaluation, and insights from the wavelet ablations. We add USFM comparisons, three-seed SDs, and independent diffusion/FAM controls. WDFNet exceeds USFM600 by 5.23/5.37 points in Accuracy/F1, and its Accuracy gain over AFG+MLP reaches 7.71 points under speckle.

1. USFM comparison

Thank you for suggesting USFM. Using its released weights and official fine-tuning recipe, we obtain:

ModelTraining epochsAccuracy (%)Macro F1 (%)
USFM20090.2489.23
USFM60090.7790.13
WDFNet60096.0095.50

These comparisons use the same PMG2000 split and preprocessing. At 600 epochs, WDFNet leads by 5.23 points in Accuracy and 5.37 in macro F1. The advantage remains after extending USFM fine-tuning from 200 to 600 epochs. Under severe speckle and matched Gaussian perturbations, Accuracy is 93.20%/91.40%, compared with 87.34%/84.17% for USFM600, supporting the benefit under image perturbations as well.

2. Repeated runs and statistical evidence

Thank you for highlighting the single-run reporting beyond PMG2000. We extended the repeated-run evaluation to the other five datasets, conducting three independent training runs to quantify run-to-run variability. We report the mean ± sample standard deviation of Accuracy and macro F1 in percentage points.

DatasetAccuracy mean ± SD (%)Macro F1 mean ± SD (%)
PMG200096.00 ± 0.1595.50 ± 0.13
UCPM220093.76 ± 0.6193.31± 0.79
BUSI92.27 ± 0.8990.98 ± 0.95
Chest-CT92.17 ± 1.3291.74 ± 1.63
Skin-Dermoscopic92.37 ± 1.6992.08 ± 1.02
Brain-MRI82.32 ± 1.2581.78 ± 1.43

Against USFM600, the paired image-bootstrap 95% CIs are [1.59, 8.99] for Accuracy and [1.39, 9.37] for macro F1 differences.

3. AFG + MLP without diffusion

Following your suggestion, we train the same ConvNeXt–Haar–AFG image architecture with an MLP and end-to-end cross entropy.

Prediction pathwayClean Acc / macro F1 (%)Speckle Acc / macro F1 (%)
AFG + end-to-end MLP93.14 / 92.5385.49 / 84.52
WDFNet96.00 / 95.5093.20 / 91.83

These comparisons use the paper's training settings. WDFNet's Accuracy/F1 gains are 2.86/2.97 points on clean images and 7.71/7.31 points under speckle. The larger speckle margin highlights the added prediction pathway's value when image conditions are perturbed.

4. Independent FAM ablation

Thank you for suggesting this control. We replace FAM with fixed addition, ht=Vg+zth_t=V_g+z_t, keeping the other components.

BUSI conditioningAccuracy (%)Macro F1 (%)
Fixed addition (w/o FAM)89.90 ± 1.0788.62 ± 1.66
FAM (WDFNet)92.16 ± 1.8090.23 ± 1.25

Values are mean ± SD over five training seeds, using the paper's training settings and original BUSI split. FAM raises mean Accuracy/F1 by 2.26/1.61 points.

5. Contribution and repeatability

Our focus is adaptive frequency conditioning during label denoising: AFG selects image details, and FAM adapts the condition to the label state. Paired five-fold BUSI evaluation gives FAM 90.77% Accuracy / 89.59% macro F1, versus 88.33% / 86.43% with fixed addition. Inner validation selects the training duration before refitting on all four training folds. The higher fold means support FAM's contribution across data partitions.

We will add the USFM and AFG+MLP comparisons, together with the repeated-run SD summary, to the main text.

Thank you again for your valuable suggestions. We have now uploaded the revised manuscript, incorporating the USFM comparison (Section 4.2), independent FAM ablation and five-fold cross-validation (Section 4.3 and Table 4), and the comparison with AFG+MLP under clean and perturbed image conditions (Section 4.4 and Table 5). Additional statistical results are provided in our rebuttal. We appreciate your feedback and hope the revised manuscript and additional experiments address your concerns.

Reviewer_9wdrOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The primary contribution is a pair of conditioning modules for label diffusion: AFG implements semantic-guided frequency selection, and FAM implements state-dependent conditioning alignment. The primary novelty is adaptive frequency conditioning within label-space diffusion, rather than frequency modeling alone. This paper deals only with image pixels (i.e. with spatial frequencies in the reconstructed image); this paper does not deal with frequencies or wavelets in the raw (beam formed) acoustic waveforms that form the basis of ultrasound images.

Critical strengths

This is an overall well written paper, and the authors seem to have done a good job defining the scope of their novelty and results, without over-claiming or overstating. Most of the ablations are thorough, and multiple datasets are used. The appendix provides many additional details for reproducibility.

Critical weaknesses

There are two main areas for improvement. First, the otherwise thorough ablations should better isolate the contributions of FAM. What happens if FAM is removed entirely, and replaced with some reasonable static heuristics? Ideally I'd like to see an ablation that traces each contribution of FAM, e.g. explore the continuum from FAM all the way down to replacing replacing scale and shift vectors with basic cross attention and replacing FAM with basic Feature-wise Linear Modulation (FiLM) by Perez et al., 2018. Second, most results lack statistical significance testing. P-values (or confidence intervals or similar) should be reported where possible to clarify which improvements are statistically significant. Additional (less critical) areas for improvement would be doing cross-fold validation on the smaller datasets, and clarifying up front that this paper concerns spatial frequencies in image pixels, not acoustic frequencies in ultrasound's raw waveforms.

Questions for authors

Please report confidence intervals and/or paired significance tests for the primary results in this paper. Can you better ablate the individual contributions of FAM in time for your rebuttal or for future work?

7 comments

Thank you for recognizing the clear presentation, careful scope, thorough ablations, and reproducibility details. Below we report statistical tests, FAM replacement controls and BUSI five-fold validation, and clarify spatial-frequency terminology.

1. Statistical evidence

Thank you for suggesting uncertainty estimates. We now report approximate t-based 95% confidence intervals for the mean:

MethodAccuracy mean ± SD (%)Macro F1 mean ± SD (%)Accuracy CI (%)Macro F1 CI (%)
DiffMIC-v293.40 ± 0.1693.20 ± 0.17[93.00, 93.80][92.78, 93.62]
WDFNet96.00 ± 0.1595.50 ± 0.13[95.63, 96.37][95.18, 95.82]

In the additional PMG2000 comparison, both WDFNet and USFM train for 600 epochs. WDFNet gains 5.23 points in Accuracy and 5.37 in macro F1, with paired image-bootstrap 95% CIs of [1.59, 8.99] and [1.39, 9.37]. Both difference intervals favor WDFNet.

2. FAM ablation

Following your suggestion, we replace FAM with fixed addition, ht=Vg+zth_t=V_g+z_t, while retaining AFG and the noisy-label/timestep/prior inputs. FAM generates FiLM-style scale and shift from the label state; fixed addition tests a static replacement.

BUSI conditioningAccuracy (%)Macro F1 (%)
Fixed addition (w/o FAM)89.90 ± 1.0788.62 ± 1.66
FAM (WDFNet)92.16 ± 1.8090.23 ± 1.25

Values are mean ± SD over five training seeds, using the paper's training settings and original BUSI split.

We also evaluated linear-generator FiLM and cross-attention with a similar parameter count on BUSI. Linear FiLM retains the state inputs and parameter count, using a linear scale/shift generator.

ConditioningAccuracy (%)Macro F1 (%)
FAM / WDFNet92.1690.23
Fixed addition88.6686.11
Linear FiLM91.2489.70
Cross-attention89.1886.51

3. Meaning of frequency

Thank you for highlighting this distinction. Frequency means spatial variation in B-mode image pixels; Haar operates on the same image as the semantic encoder.

We will make this definition explicit in the revised paper.

4. Cross-validation

Thank you for suggesting evaluation across splits. We completed paired five-fold BUSI evaluation:

ConditioningAccuracy (%)Macro F1 (%)
Fixed addition88.33 ± 1.2386.43 ± 2.33
FAM (WDFNet)90.77 ± 1.9589.59 ± 2.08

Values are fold mean ± SD. Inner validation selects the training duration; each model is then retrained on all four training folds before testing on the held-out fold. FAM improves average Accuracy/F1 by 2.44/3.16 points, supporting its contribution across train/test partitions.

We will add the independent FAM controls and five-fold summary to the revised main text, with statistical details in the appendix.

Reviewer_9wdrOpenReview ↗

Improved ablations, however cross-validation shows overlapping confidence intervals

Thank you for the new ablations. It's great to see statistically significant (albeit small) improvement in your rebuttal sections 1. Statistical evidence and 2. FAM ablation. Thank you also for adding cross-validation. It is encouraging to see improvements generally hold up during CV, but I'm a little concerned about the large SD margins. Fixed addition 88.33+1.23 = 89.56, and FAM (WDFNet) 90.77 - 1.95 = 88.82; these means are separated by a single SD from a single method (e.g. 89.56 < 90.77 and 88.33 < 88.82), but their overlapping single-SD ranges may be problematic: 89.56 is not < 88.82.

Further, these CV results only compare to fixed addition. The FAM ablation showed that the strongest contender is Linear FiLM.

Please expand the 5-fold CV to include comparison with Linear FiLM. Please also provide detailed results for each of the 5 folds. It will be important to see how many of the 5 folds have statistically significant improvement of FAM over Linear FiLM.

Thank you for your constructive follow-up and for recognizing the additional evidence.

We agree that extending the five-fold comparison to Linear FiLM, the strongest alternative in our conditioning ablation, would provide a more informative evaluation.

Regarding the overlapping SD ranges, the reported SDs reflect variability across data partitions rather than uncertainty in the paired performance differences. Since both methods are evaluated on identical folds, we believe that fold-wise paired comparisons and confidence intervals for the performance differences would better address this concern.

We are extending the BUSI five-fold evaluation to Linear FiLM using the same partitions and training protocol. We will report the individual fold results, paired performance differences, and statistical tests on the held-out predictions as soon as the evaluation is completed.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.