acceptodds
Under review as a conference paper at ICLR 2027

Learning from K-Space: Acquisition-Informed Pretraining for Label-Efficient Segmentation from Undersampled 3D MRI

Abstract

Accurate three-dimensional (3D) brain tumor segmentation from MRI faces two practical challenges: costly voxel-wise annotations and accelerated acquisition through undersampling. Existing MRI segmentation paradigms either operate on reconstructed images without direct access to acquisition-domain information or incorporate k-space as an additional input without explicitly modeling its task-relevant structure, which can degrade segmentation performance under limited supervision. To address this, we propose an acquisition-informed self-supervised learning framework for effective label-efficient 3D brain tumor segmentation. Specifically, our network learns directly from undersampled k-space together with the corresponding zero-filled reconstructed image, combining acquisition-domain frequency information with essential spatial context without requiring a high-quality MRI reconstruction. To utilize k-space more effectively, our empirical analysis reveals structured pathology-associated spectral deviations that exhibit meaningful spatial correspondence with tumor regions. Consequently, we formulate these spectral residuals as a soft spectral prior during pretraining to guide representation learning. Finally, we introduce an auxiliary lesion-presence branch during fine-tuning that softly modulates dense predictions to effectively suppress unsupported false positives. Extensive experiments on multimodal brain tumor MRI demonstrate consistent improvements under limited supervision, highlighting the value of pathology-aware \(k\)-space learning for label-efficient segmentation from accelerated MRI.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
K-space representation learning, self-supervised learning, frequency-domain learning, semantic segmentation, undersampled MRI, representation transfer

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_3wosOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes K2S, a label-efficient framework for brain tumor segmentation from undersampled multimodal MRI. The model jointly encodes undersampled k-space and its zero-filled reconstruction. During pretraining, it combines image reconstruction with prediction of an annotation-free spectral prior derived from deviations in the radial k-space energy profile. During fine-tuning, a slice-level lesion-presence branch softly gates TC and ET predictions to suppress false positives. Experiments are conducted on image-derived synthetic k-space from BraTS 2024, using 100 subjects for pretraining, 50 for fine-tuning, 20 for validation, and 20 for testing. The proposed method outperforms several image-domain and pretraining-based baselines and includes component and label-efficiency studies.

Critical strengths
  1. Task-oriented frequency-domain pretraining. The spectral-prior objective is an interesting attempt to convert an annotation-free k-space statistic into explicit supervision for segmentation-oriented representation learning. The ablation suggests that it improves positive-case Dice and HD95 over reconstruction-only pretraining.
  2. Effective false-positive suppression. The presence head and soft gating directly address a practical failure mode under limited supervision. They substantially improve TC/ET reliability and contribute to the strong overall results.
Critical weaknesses
  1. The central “pathology-specific” and “acquisition-informed” claims are not sufficiently established. The k-space data are synthesized by applying a slice-wise 2D FFT to already reconstructed and preprocessed BraTS images, rather than obtained from raw MRI acquisitions. The experiments therefore exclude coil sensitivities, native phase, acquisition noise, and scanner-dependent effects. Moreover, the zero-filled image is deterministically obtained from the same undersampled k-space, so the results primarily demonstrate the value of dual-domain representations and auxiliary supervision, rather than access to additional acquisition information. The proposed residual also lacks strong null controls. Its GT-free localization precision is only 0.340, and the paper does not compare against random frequency bands, generic high-pass targets, shuffled priors, or conventional edge responses. It is therefore unclear whether the residual captures pathology-specific structure rather than generic contrast or anatomical boundaries.

  2. The current ablations and evaluation do not cleanly isolate the source or robustness of the gains. The ablation is cumulative rather than factorial: there is no presence-only variant, no strong baseline equipped with the same presence mechanism, and no parameter-matched dual-domain model without the proposed spectral target. Thus, the contribution of the spectral prior cannot be separated from model capacity, dual-domain processing, or presence supervision. The evaluation further relies on a single 20-subject test split, one acceleration factor, and one mask protocol, without repeated runs, confidence intervals, or statistical tests. These limitations weaken both attribution and robustness claims.

Questions for authors
  1. Which data split was used for the spectral analysis, GT-guided oracle, modality selection, and hyperparameter choices in Section 3? Were test-set images or annotations used at any stage?
  2. What exactly is provided to the image encoder: complex-valued data, real and imaginary channels, or magnitude images? Since the zero-filled image is obtained by IFFT from the same k-space, what information is available only to the k-space branch?
  3. How is “Absent FP” defined, and how many absent TC and ET cases are included in the test set?
  4. Can the authors provide null controls, including random-band selection, generic Fourier-energy or high-pass targets, shuffled priors, and sensitivity to the number of selected bands and Savitzky–Golay parameters?
  5. Can the authors report a factorial ablation including reconstruction plus presence supervision, spectral pretraining without presence supervision, and a strong parameter-matched baseline with the same presence head?
9 comments
AuthorsOpenReview ↗

Dual-Domain Clarifications, Null Controls and Statistical Analysis

Thank you for your thorough review and detailed assessment. Below, we clarify our dual-domain formulation, experimental setup, and expanded ablations.

Response to W1(a). Clarification of acquisition-informed learning

We acknowledge that our k-space is synthetically generated from preprocessed BraTS images and does not capture real acquisition effects such as coil sensitivities, native phase, and scanner noise. Accordingly, we do not claim access to additional independent acquisition information through dual-domain inputs; rather, they provide complementary representations with different inductive biases. Importantly, our contribution lies in explicitly modeling lesion-associated spectral deviations rather than simply adding Fourier representations. As shown in Table R5, residual-based spectral supervision outperforms raw Fourier-spectrum supervision in downstream segmentation, supporting the benefit of our proposed spectral prior. We have clarified these limitations and moderated our claims accordingly, while real multi-coil raw k-space validation remains an important direction for future work.

Table R5. Full 2D pretraining target comparison after fine-tuning.

Pretraining TargetWT DiceTC DiceET DiceAvg. Subject DiceAvg. Nonempty Dice
Raw Fourier spectrum0.7150.4870.4460.5490.535
Spectral residual (Ours)0.7270.5590.5180.6010.563

Response to W1(b) & Q4(a). Spectral Null Controls: Localization and Downstream Segmentation

We clarify that our pretraining uses the full spectral residual distribution as the supervision target, while Top-5 band selection is used only for preliminary localization analysis. We evaluate the null controls at two levels:

  1. Localization comparison (Table R2): For frequency-band selection methods, we compare Top-5 selected bands under the same localization protocol. Our residual-based selection achieves higher lesion localization rates than random-band, high-pass, and Sobel-edge controls on an independent 40-subject cohort.

  2. Pretraining comparison (Table R5): Unlike band-selection methods, the raw Fourier spectrum does not explicitly select frequency bands. We therefore compare it as an alternative pretraining target by replacing the spectral residual distribution with the raw Fourier-spectrum distribution, followed by identical pretraining and fine-tuning. Our spectral residual consistently outperforms raw-spectrum supervision across all reported Dice metrics, demonstrating its effectiveness beyond generic frequency-domain supervision.

Table R2. Component-hit rate on an independent cohort.

Target / ModalityGT OracleResidual (Ours)High-pass bandsSobel edgeRandom bands
WT / FLAIR0.8560.8000.2440.2060.504
Edema / FLAIR0.8560.7560.1750.1380.442
ET / T1ce0.7550.5590.2160.2350.491
TC / T1ce0.7650.5590.2060.2450.504

Response to Q4(b). Parameter sensitivity analysis

As suggested, we evaluated the parameter sensitivity of the spectral residual in Table R7. Localization hit rates remain stable across different numbers of selected bands (K = 3 to 10), Savitzky-Golay window sizes (31 to 71), and polynomial orders (1 to 3).

AuthorsOpenReview ↗

Statistical Analysis (Continued) and Factorial Ablation

Table R7. Sensitivity analysis of the spectral residual on the independent 40-subject cohort (component-hit rate).

Parameter SettingWT / FLAIREdema / FLAIRET / T1ceTC / T1ce
Number of Selected Bands (Top-K)
K = 30.7870.7560.5590.569
K = 5 (Default)0.8000.7560.5590.559
K = 70.8250.7810.5590.569
K = 100.8250.7620.6080.608
Savitzky-Golay Window Size
Window = 310.7940.7620.5490.559
Window = 51 (Default)0.8000.7560.5590.559
Window = 710.8060.7620.5590.559
Savitzky-Golay Polynomial Order
Order = 10.8060.7620.5590.559
Order = 2 (Default)0.8000.7560.5590.559
Order = 30.8000.7690.5390.549

Response to W2 (a) & Q5. Factorial ablation and parameter matching

Table R1 (also updated as Table 2 in the revised manuscript) includes the previously missing reconstruction + presence + gating control and confirms that our full model outperforms this nearly parameter-matched non-spectral control (48.04M vs. 48.05M parameters). These results demonstrate the independent contribution of spectral supervision beyond model capacity and presence-aware adaptation.

Table R1. Expanded ablation study on the test set (updated as Table 2 in the revised manuscript). All variants share the same reconstruction pretraining. Absent FP denotes the fraction of lesion-absent cases (8 for TC, 8 for ET) with false-positive predictions.

VariantSpectral PTSpectral FT/InfPresenceSoft GatingAvg. Dice ↑NE Dice ↑NE HD95 (mm) ↓Absent FP (TC / ET) ↓
Without Spectral Prior (Controls)
Recon-only baseline✗✗✗✗0.5440.52530.244/8, 4/8
+ Presence head✗✗✓✗0.5400.51926.654/8, 4/8
+ Presence + Gating (Control)✗✗✓✓0.5830.53624.873/8, 3/8
With Spectral Prior
+ Spectral PT✓✗✗✗0.4280.55321.348/8, 8/8
+ Spectral PT + Presence✓✗✓✗0.5020.53918.016/8, 5/8
+ Spectral PT + Presence + Gating✓✗✓✓0.5860.54017.943/8, 3/8
+ Spectral PT & FT/Inf + Presence✓✓✓✗0.5620.55316.624/8, 4/8
+ Spectral PT & FT/Inf + Presence + Gating (Ours)✓✓✓✓0.6010.56312.693/8, 3/8

AuthorsOpenReview ↗

Acquisition robustness and Statistical Reliability

Response to W2(b). Acquisition robustness

Thank you for this suggestion, which helped us enrich the paper with a broader robustness evaluation. Our proposed method consistently ranks first across different acceleration factors and different sampling trajectories. Evaluated across different acceleration factors (R=2, 4, 6, and 8) as well as four sampling patterns without retraining (Table R4), our method achieves the highest average subject Dice in every setting. We have also added a radar plot in Figure 5 (Fig. 5b) of the revised manuscript illustrating average subject Dice across these various undersampled settings.

Table R4. Average subject Dice across seven retrospective acquisition settings on the 20-subject test set.

MethodCart. R=2Cart. R=4Cart. R=6Cart. R=82D Gauss. R=42D Poisson R=42D Radial R=4
3D U-Net0.4370.4060.4610.3730.4890.4430.382
nnUNet0.4910.4870.4970.5210.4390.4540.470
MedNeXt0.5080.5190.5160.4010.4580.4400.428
SwinUNETR0.4390.4230.3910.4190.4320.4400.209
AMAES (U-Net-XL)0.4810.4540.4000.3800.4790.5000.341
AMAES (MedNeXt-L)0.4940.4280.4520.4410.5170.5130.217
KMAE0.4020.3620.2880.1850.3210.3290.393
Ours0.6350.6010.5600.5870.6260.6130.480

Response to W2(c). Statistical Reliability

Following your suggestion, we have added the paired bootstrap 95% confidence intervals (Table R6). Paired bootstrap 95% confidence intervals confirm consistent subject-level gains across all baselines. Using subject-level paired bootstrap resampling (20,000 iterations, n=20), the 95% confidence intervals (CIs) for average subject Dice improvement exclude zero against all seven competing baselines.

Table R6. Paired differences (Ours - baseline) with 95% bootstrap CIs on the 20-subject test set. Positive Dice and negative HD95 favor Ours.

Baseline ComparisonΔ Avg. Subject Dice ↑Δ Avg. NE Dice ↑Δ Avg. NE HD95 (mm) ↓
MedNeXt+0.082 [0.021, 0.168]+0.069 [0.015, 0.138]-9.12 [-21.81, -0.77]
nnUNet+0.115 [0.030, 0.215]+0.019 [-0.005, 0.046]+0.37 [-1.42, 2.34]
AMAES (U-Net-XL)+0.148 [0.067, 0.246]+0.094 [0.052, 0.140]-7.82 [-18.98, 0.53]
AMAES (MedNeXt-L)+0.173 [0.071, 0.298]+0.053 [0.035, 0.073]-9.09 [-19.76, -0.34]
SwinUNETR+0.179 [0.058, 0.314]+0.016 [-0.018, 0.053]-13.28 [-27.97, -0.73]
3D U-Net+0.196 [0.090, 0.318]+0.089 [0.046, 0.136]-11.55 [-23.60, -3.15]
KMAE (frozen + matched nnUNet)+0.239 [0.123, 0.366]+0.101 [0.050, 0.158]-28.91 [-42.92, -15.65]

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

This paper proposes a label-efficient framework for 3D brain tumor segmentation from undersampled multimodal MRI. Specifically, it explicitly exploits pathology-related spectral structure in k-space: during pretraining, the model jointly learns from undersampled k-space and zero-filled images using reconstruction and a spectral-prior prediction objective; during fine-tuning, a slice-level presence branch softly gates tumor predictions to reduce false positives. Experiments on BraTS 2024 show improved Dice and HD95 over several end-to-end and pretraining baselines, with additional ablations and label-efficiency studies.

Critical strengths
  1. The paper aims to address two limitations of the existing brain tumor segmentation task. The motivation is good, and the central premise of pathology-related k-space structure was carefully validated. Then the paper uses spectral residuals as a soft prior during pretraining, which provides a meaningful acquisition-specific contribution beyond standard dual-domain learning.

  2. Overall, the performance in the limited-label setting is promising. The proposed framework achieves the strongest overall results among the evaluated end-to-end and pretraining baselines, with especially notable gains for the more difficult TC and ET regions, and it remains competitive across 10, 25, and 50 labeled subjects. This supports the practical relevance of the approach for label-efficient.

Critical weaknesses
  1. The claim regarding performance in the acquisition domain is validated only on synthetic k-space data. Further validation using real acquired data would strengthen the claim and better demonstrate the practical applicability of the proposed method.

  2. The ablation study shows that adding the spectral prior to the reconstruction-only model decreases the average subject-level Dice score from 0.544 to 0.428, while the subsequent introduction of the presence head and soft gating recovers and ultimately improves performance. Given this behavior, the necessity and individual contribution of each component within the framework remain unclear and warrant further justification.

Questions for authors

Q1. Can the authors provide the missing ablation with reconstruction pretraining + presence head + soft gating, but without the spectral-prior objective?

Q2. How robust are the reported gains to different acquisition settings, such as other acceleration factors, sampling masks, or k-space noise levels? The current experiments appear to use only one retrospective configuration (R=4R=4, Cartesian undersampling, 8% central lines).

Q3. Is it possible to evaluate on a dataset with raw k-space (rather than synthesized k-spece as in BraTS)?

4 comments
AuthorsOpenReview ↗

Expanded Ablations and Cross-Protocol Acquisition Robustness

We thank the reviewer for the positive assessment and the suggestions on acquisition robustness and component attribution.

Response to W2 & Q1. Necessity of individual components

The requested reconstruction + presence + gating control is now included (Table R1, updated as Table 2 in the revised manuscript). It confirms that the full method’s benefit cannot be attributed solely to the presence mechanism: with nearly identical capacity and presence-aware adaptation, spectral pretraining improves positive-case boundary accuracy, while the complete spectral pathway further improves nonempty Dice and HD95. The decrease in subject Dice from spectral pretraining alone reflects additional false positives on absent cases; presence-aware adaptation compensates for this limitation.

Table R1. Expanded ablation study on the test set (updated as Table 2 in the revised manuscript). All variants share the same reconstruction pretraining. Absent FP denotes the fraction of lesion-absent cases (8 for TC, 8 for ET) with false-positive predictions.

VariantSpectral PTSpectral FT/InfPresenceSoft GatingAvg. Dice ↑NE Dice ↑NE HD95 (mm) ↓Absent FP (TC / ET) ↓
Without Spectral Prior (Controls)
Recon-only baseline✗✗✗✗0.5440.52530.244/8, 4/8
+ Presence head✗✗✓✗0.5400.51926.654/8, 4/8
+ Presence + Gating (Control)✗✗✓✓0.5830.53624.873/8, 3/8
With Spectral Prior
+ Spectral PT✓✗✗✗0.4280.55321.348/8, 8/8
+ Spectral PT + Presence✓✗✓✗0.5020.53918.016/8, 5/8
+ Spectral PT + Presence + Gating✓✗✓✓0.5860.54017.943/8, 3/8
+ Spectral PT & FT/Inf + Presence✓✓✓✗0.5620.55316.624/8, 4/8
+ Spectral PT & FT/Inf + Presence + Gating (Ours)✓✓✓✓0.6010.56312.693/8, 3/8

Response to Q2. Robustness across acquisition settings

Thank you for this helpful suggestion, which helped us to further strengthen and enrich the paper. Our proposed method achieves the highest average subject Dice among the evaluated methods in all seven retrospective sampling settings (Table R4), including Cartesian R=2/4/6/8 and three alternative R=4 sampling patterns. These results show that the relative performance advantage is not confined to the original Cartesian R=4 mask. The experiments use cross-protocol inference without retraining. To visually summarize this robustness across various undersampled settings, we have also added a radar plot of average subject Dice in Figure 5 (Fig. 5b) of the revised manuscript.

Response to W1 & Q3. Validation on raw k-space data

We agree that evaluation on scanner-acquired multi-coil raw k-space would strengthen clinical validation. However, to our knowledge, no publicly available dataset currently combines raw k-space with 3D multimodal brain tumor MRI and voxel-wise annotations. BraTS provides expert 3D annotations but only on reconstructed images. We have acknowledged this limitation in the revised Conclusion and plan to collect clinical raw k-space data in future work.

AuthorsOpenReview ↗

Cross-Protocol Acquisition Robustness (Continued)

Table R4. Average subject Dice across seven retrospective acquisition settings on the 20-subject test set.

MethodCart. R=2Cart. R=4Cart. R=6Cart. R=82D Gauss. R=42D Poisson R=42D Radial R=4
3D U-Net0.4370.4060.4610.3730.4890.4430.382
nnUNet0.4910.4870.4970.5210.4390.4540.470
MedNeXt0.5080.5190.5160.4010.4580.4400.428
SwinUNETR0.4390.4230.3910.4190.4320.4400.209
AMAES (U-Net-XL)0.4810.4540.4000.3800.4790.5000.341
AMAES (MedNeXt-L)0.4940.4280.4520.4410.5170.5130.217
KMAE0.4020.3620.2880.1850.3210.3290.393
Ours0.6350.6010.5600.5870.6260.6130.480

Dear Reviewer dDQL,

Thank you again for your positive assessment and helpful suggestions. We would like to briefly summarize the updates made in response to your comments:

  • Cross-protocol acquisition robustness: Without any fine-tuning, our method achieves the highest average subject Dice across all seven retrospective undersampling settings covering Cartesian R=2, 4, 6, 8 and 2D Gaussian, Poisson-disk, and Radial trajectories (Table R4, Revised Figure 5).
  • Component necessity and non-spectral control: We added the requested reconstruction + presence + gating control (Table R1, Revised Table 2), confirming that neither the spectral prior nor presence gating alone achieves the full model performance.
  • Raw k-space dataset scope: We explicitly noted the lack of annotated 3D multimodal raw k-space brain tumor datasets and our plan for clinical multi-coil validation in the revised Conclusion.

We welcome any remaining questions and thank you again for your supportive review.

Dear Reviewer dDQL,

If you have any further questions, concerns, or suggestions, please do not hesitate to reach out. We would be delighted to engage in a constructive discussion to enhance the quality and impact of our work.

Thank you once again for your time and thoughtful review.

Reviewer_xYjpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses label-efficient 3D brain tumor segmentation from undersampled multimodal MRI. The authors propose K2S, an acquisition-informed framework that jointly processes undersampled k-space and its corresponding zero-filled reconstruction through a dual-domain architecture.

The central idea is to exploit spectral residuals derived from radial k-space energy distributions as an explicit self-supervised pretraining target. During pretraining, the model jointly optimizes image reconstruction and spectral-prior prediction to learn anatomical and frequency-domain representations. During fine-tuning, an auxiliary slice-level lesion-presence branch is introduced to softly modulate voxel-wise tumor predictions, aiming to reduce false positives.

The method is evaluated on BraTS 2024 using retrospectively simulated fourfold Cartesian undersampling, with 100 subjects for pretraining and up to 50 labeled subjects for fine-tuning. The authors report improvements in average subject-level Dice and nonempty-case HD95 over the evaluated segmentation and pretraining-based baselines, supported by ablation and label-efficiency experiments.

Critical strengths
  1. Meaningful problem and method design.

The paper addresses the practical challenges of undersampled MRI and limited segmentation annotations. Its use of spectral residuals as an explicit self-supervised pretraining target provides an interesting extension of existing k-space representation learning approaches.

  1. Promising empirical results.

Table 1 reports improvements in average subject-level Dice (0.601 vs. 0.519 for the strongest end-to-end baseline) and average nonempty HD95 (12.693 vs. 16.924).

Table 2 further provides useful evidence that spectral-guided pretraining improves positive-case segmentation, while presence-aware adaptation helps suppress false positives.

These findings suggest that the overall framework has practical potential.

Critical weaknesses
  1. Insufficient validation of the spectral prior's scientific motivation.

Although the authors report lesion-associated spectral responses, the residuals may also reflect normal anatomical structures and other non-pathological frequency variations.

Moreover, the spatial precision metric in Eq. (3) differs from conventional voxel-level precision. The evidence does not sufficiently establish the pathology specificity or unique advantage of the proposed spectral prior over simpler frequency-based alternatives.

This limitation directly affects the strength of the scientific motivation underlying the central contribution.

  1. The independent contribution of spectral supervision remains unclear.

Table 2 shows that adding the spectral prior improves nonempty Dice and HD95 but substantially reduces overall subject-level Dice.

While presence-aware adaptation subsequently improves performance, a critical ablation combining reconstruction pretraining with presence-aware adaptation and soft gating, but without spectral supervision, is missing.

Therefore, the extent to which the final improvements depend on the proposed spectral prior rather than the subsequent presence-aware mechanism remains uncertain.

This missing comparison limits the evidence supporting the paper's central novelty claim.

Questions for authors
  1. Independent contribution of the spectral prior.

Could the authors report results for reconstruction pretraining combined with the presence head and soft gating, but without spectral supervision?

This would help isolate the spectral prior's contribution to the final performance in Table 2.

  1. Validity and specificity of spectral residuals.

How do the authors distinguish pathology-related spectral residuals from variations caused by normal anatomy or image characteristics?

  1. KMAE baseline adaptation.

Why is the pretrained KMAE reconstruction network frozen while its segmentation head is replaced with a full 3D nnUNet?

Have the authors evaluated end-to-end fine-tuning or other adaptation strategies to establish the robustness of this comparison?

4 comments
AuthorsOpenReview ↗

Spectral Prior Validation, Expanded Ablations, and Clarifications on KMAE Finetuning

Thank you for your thoughtful review and constructive feedback. Below, we provide new ablation and control experiments along with clarifications on our spectral prior.

Response to W1(a) & Q2: Validity and specificity of spectral residuals.

We clarify that the spectral prior is not designed to distinguish tumors from normal tissue, but to localize potentially abnormal regions where tumors are expected to reside. We acknowledge that normal anatomical structures may also exhibit spectral variations, leading to false positives. Our ablation results in Table R1 demonstrate that the spectral prior improves non-empty lesion Dice from 0.525 to 0.553, despite increased false positives (4/8 to 8/8). Importantly, the presence-aware adaptation reduces false positives to 3/8, while the full model further improves lesion Dice to 0.563, demonstrating the complementary roles of spectral localization and presence-aware suppression.

Table R1. Expanded ablation study on the test set (updated as Table 2 in the revised manuscript). All variants share the same reconstruction pretraining. Absent FP denotes the fraction of lesion-absent cases (8 for TC, 8 for ET) with false-positive predictions.

VariantSpectral PTSpectral FT/InfPresenceSoft GatingAvg. Dice ↑NE Dice ↑NE HD95 (mm) ↓Absent FP (TC / ET) ↓
Without Spectral Prior
Recon-only baseline✗✗✗✗0.5440.52530.244/8, 4/8
+ Presence head✗✗✓✗0.5400.51926.654/8, 4/8
+ Presence + Gating✗✗✓✓0.5830.53624.873/8, 3/8
With Spectral Prior
+ Spectral PT✓✗✗✗0.4280.55321.348/8, 8/8
+ Spectral PT + Presence✓✗✓✗0.5020.53918.016/8, 5/8
+ Spectral PT + Presence + Gating✓✗✓✓0.5860.54017.943/8, 3/8
+ Spectral PT & FT/Inf + Presence✓✓✓✗0.5620.55316.624/8, 4/8
+ Spectral PT & FT/Inf + Presence + Gating (Ours)✓✓✓✓0.6010.56312.693/8, 3/8

Response to W1(b): Spatial precision metric

We clarify that the spatial precision in Eq. (3) is not intended to measure voxel-level segmentation accuracy, but rather the coarse localization relevance of spectral responses. Specifically, it measures the proportion of response areas that belong to components spatially corresponding to the lesion, based on a predefined center/overlap criterion. Our goal is to assess whether spectral responses provide useful coarse localization cues, rather than precise tumor delineation.

Response to W1(c): Comparison with simpler frequency-based alternatives

On an independent 40-subject cohort, residual-based selection consistently achieves higher localization hit rates than random-band, high-pass, and Sobel-edge controls across all four tumor regions (Table R2). This supports its lesion association beyond generic frequency selection or anatomical edge responses.

Table R2. Component-hit rate on an independent cohort.

Target / ModalityGT OracleResidual (Ours)High-pass bandsSobel edgeRandom bands
WT / FLAIR0.8560.8000.2440.2060.504
Edema / FLAIR0.8560.7560.1750.1380.442
ET / T1ce0.7550.5590.2160.2350.491
TC / T1ce0.7650.5590.2060.2450.504

AuthorsOpenReview ↗

Spectral Prior Validation, Expanded Ablations, and Clarifications on KMAE Finetuning (Continued)

Response to W2 & Q1. Independent contribution of spectral supervision

The expanded ablation isolates a benefit from spectral supervision beyond presence-aware adaptation (Table R1, updated as Table 2 in the revised manuscript). With presence supervision and soft gating fixed, spectral pretraining reduces nonempty HD95 by 27.8% relative to the non-spectral control; the complete model reduces it by 48.9%. Spectral pretraining alone improves positive-case Dice and HD95 but increases absent-case false positives, explaining its lower subject-level Dice. The presence head and soft gating supply complementary lesion-existence evidence and suppress unsupported predictions. Thus, the spectral and presence-aware components play complementary roles in balancing sensitivity and specificity.

Response to Q3. KMAE baseline adaptation

We clarify that KMAE was originally designed for 2D cardiac MRI with a lightweight segmentation head, which lacks decoding capacity for 3D multimodal brain tumors. To isolate representation quality, we equipped the pretrained KMAE encoder with the exact same 3D nnUNet decoder used in Ours. This freezing strategy is consistent with our own downstream adaptation protocol, where the pretrained encoders are also frozen. We now additionally evaluate end-to-end fine-tuning of its pretrained encoder. Our advantage over KMAE persists under both frozen-encoder and end-to-end adaptation (Table R3).

Table R3. Comparison with KMAE under different adaptation protocols.

Method / protocolWT Dice ↑TC Dice ↑ET Dice ↑Avg. Subject Dice ↑Avg. NE Dice ↑Avg. NE HD95 (mm) ↓
KMAE (frozen encoder + 3D nnUNet)0.6350.2430.2070.3620.46250.462
KMAE (end-to-end fine-tuning + 3D nnUNet)0.6300.2200.1800.3430.43247.775
Ours0.7270.5590.5180.6010.56312.693

Dear Reviewer xYjp,

Thank you again for your constructive feedback. As the discussion period is underway, we would like to briefly highlight the new experiments and revisions addressing your main concerns:

  • Spectral prior validity and controls: New evaluations on an independent 40-subject cohort confirm stronger lesion localization than random-band, high-pass, and Sobel-edge controls (Table R2), and full residual supervision outperforms raw Fourier-spectrum supervision in downstream segmentation (Table R5).
  • Independent contribution of spectral supervision: Expanded ablations including the non-spectral control (Table R1, Revised Table 2) confirm that the proposed spectral supervision substantially improves lesion boundary accuracy (reducing nonempty HD95 by 48.9%), while presence gating suppresses false positives.
  • KMAE end-to-end fine-tuning: We added end-to-end fine-tuning results for KMAE with the matched 3D nnUNet decoder (Table R3), showing that our performance advantage persists under both frozen and end-to-end adaptation.

We would greatly appreciate your thoughts on these additions and welcome any remaining questions. If these updates address your concerns, we would be grateful if you could reconsider your assessment.

Thank you for your time and consideration.

Dear Reviewer xYjp,

If you have any further questions, concerns, or suggestions, please do not hesitate to reach out. We would be delighted to engage in a constructive discussion to enhance the quality and impact of our work.

Thank you once again for your time and thoughtful review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.