acceptodds
Withdrawn from ICLR 2027

SFCI-Mamba: Factorized Amplitude–Phase Modeling with Spatial–Frequency Interaction for Skin Lesion Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Skin lesion segmentation requires both boundary detail and global structure. We study two questions in spatial–frequency state-space modeling: how to model Fourier amplitude and phase, and how spatial and spectral features should interact before fusion. We present SFCI-Mamba, a compact encoder–decoder built around spatial-frequency collaborative state-space blocks (SFCBs). Each SFCB independently maps amplitude- and phase-domain representations with Mamba, while dual-context interaction (DCI) lets each branch retrieve from a shared spatial-frequency context using its own query before fusion. SFCI-Mamba achieves micro-averaged Dice scores of 0.9121, 0.9001, and 0.9228 on ISIC2017, ISIC2018, and PH, respectively. The PH result is obtained using the ISIC2018-trained model without fine-tuning. Among the tested variants, frequency modeling outperforms spatial replacements, joint amplitude–phase modeling gives the strongest overlap, and DCI improves over the tested interaction alternatives. These results support factorized spectral state-space modeling with pre-fusion spatial–frequency interaction for dermoscopic lesion segmentation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical Image Segmentation, State Space Models, Mamba, Spatial-Frequency Learning, Skin Lesion Segmentation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_zUsHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SFCI-Mamba, a compact U-shaped encoder-decoder for dermoscopic skin lesion segmentation. Its spatial-frequency collaborative block contains a spatial branch based on two-dimensional selective scanning and a frequency branch that applies a Fourier transform, separately maps amplitude and phase using two independent Mamba modules, reconstructs a complex spectrum, and returns it to the spatial domain. The proposed dual-context interaction module allows branch-specific queries to retrieve information from a shared spatial-frequency context before concatenation and SE-based fusion. The model is evaluated on ISIC2017 and ISIC2018, together with cross-dataset evaluation on PH2 using the ISIC2018-trained checkpoint without fine-tuning. Against eight retrained baselines, it reports Dice scores of 0.9121, 0.9001, and 0.9228, respectively, with 2.538M parameters and 0.354 GFLOPs. The ablation studies examine DCI, SE, frequency-domain modeling, and separate versus joint amplitude-phase modeling.

Critical strengths
  1. The paper specifies the dataset sizes, common training protocol, metric definitions, micro-averaging convention, thresholding rule, and handling of empty masks for HD95 and ASSD. It reports overlap, pixel-level, and boundary metrics and includes qualitative examples, failure cases, and an appropriately scoped limitations section. These choices make the empirical claims relatively easy to understand and audit.
  2. Under the reported common protocol, SFCI-Mamba obtains the best IoU and Dice among the evaluated methods on both ISIC test sets and on PH2, while remaining smaller than most conventional and state-space baselines. The component studies are also directionally consistent: the complete model outperforms the tested spatial replacements, and joint amplitude-phase modeling gives the best overlap among the evaluated spectral variants.
Critical weaknesses
  1. The methodological novelty mainly comes from combining existing components and appears relatively incremental. Spatial-frequency dual-branch modeling, Fourier amplitude-phase decomposition, visual Mamba, SE, and cross-branch attention have all been explored in related work. The main new designs are the use of independent Mamba mappings for amplitude and phase and the use of branch-specific queries over a concatenated shared context. However, the former largely replaces an existing mapping operation with state-space modules, while the latter remains structurally close to multi-head cross-attention. The paper does not provide sufficient theoretical analysis, spectral-behavior analysis, or strictly parameter-matched experiments to demonstrate that these designs introduce capabilities unavailable to existing methods. The current contribution therefore appears closer to an effective incremental architecture than a substantial methodological advance.
  2. The experiments do not establish whether the relatively small improvements are robust, and they do not fully separate the proposed mechanism from additional capacity and computation. All tables report single point estimates without multiple random seeds, confidence intervals, or statistical tests. The Dice improvements over the strongest listed baseline are 0.0038 on ISIC2017, 0.0067 on ISIC2018, and 0.0027 on PH2, while the improvement of joint amplitude-phase modeling over phase-only modeling is only 0.0015. Moreover, DCI+SE uses 2.538M parameters and 0.354 GFLOPs, compared with 2.002M parameters and 0.247 GFLOPs for no-DCI+SE. Part of the improvement may therefore result from increased model capacity or computation rather than the interaction mechanism itself.
Questions for authors
  1. Beyond combining existing modules and changing their mappings, what new modeling capability does SFCI-Mamba provide relative to prior spatial-frequency Mamba, amplitude-phase modeling, and bidirectional cross-attention approaches? Can the authors provide spectral-response or feature-behavior analyses and parameter- and FLOP-matched no-DCI, bidirectional cross-attention, and other relevant controls to show that the gains cannot be obtained from a standard dual-branch architecture or increased capacity?
  2. Can the authors report results over multiple random seeds, including means and standard deviations or confidence intervals, for the full model, the strongest baseline, and the key ablations in Tables 5-7? A paired bootstrap analysis over test images would also be informative. Please additionally specify the channel schedule, Mamba state and expansion hyperparameters, DCI pooling and upsampling procedure, normalization and rotation settings, exact split identifiers, and code-release plan.
Reviewer_zxeoOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SFCI-Mamba, a U-shaped segmentation network that combines spatial and frequency-domain state-space modeling. Its spatial-frequency collaborative blocks independently model Fourier amplitude and phase using Mamba, while a dual-context interaction (DCI) module exchanges information between spatial and frequency representations before fusion.

Critical strengths

Modeling amplitude and phase separately with Mamba provides a concrete extension of existing spatial-frequency segmentation approaches. The ablations support the usefulness of jointly modeling both components.

Critical weaknesses
  • No cross-dataset evaluations were provided, such as training on ISIC 2018 and testing on PH2. No ablation study examines how input size affects the HD95 score. Authors should conduct more meaningful ablations to demonstrate performance under stress testing.
  • Overall performance does not improve much over existing methods, despite increased model complexity, and there is no comparison with any recent papers from 2026.
  • Authors should also compare DCI against a parameter- or FLOPs-matched cross-attention alternative. This would help establish whether the improvement comes from the interaction design rather than additional capacity.
  • Authors did not provide inference latency or throughput comparisons with lighter baselines.
  • How does separate Mamba-based amplitude/phase modeling compare with a parameter-matched convolutional spectral branch? This would better isolate the contribution of state-space modeling in the frequency domain.
  • Since there are no multiple runs, it's also very difficult to prove the stability of the reported gains compared with other methods.
Questions for authors

N/A

Reviewer_6M5cOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed SFCI-Mamba for skin lesion segmentation, evaluated on ISIC 2017, ISIC 2018 and PH2. Their main motivation is to improve segmentation with better representation of boundary detail and global structure. The idea is to integrate spatial ( spatial-frequency collaborative state-space block) and frequency branch (models amplitude and phase). Before fusion, a Dual-context interaction (DCI) module is used to enable information exchange between the two branches.

The contribution is to introduce SFCI-Mamba, with two branches and DCI. The method is evaluated on one domain, skin lesions publicly available datasets. The results showed marginal improvement for the majority of the metrics, with lower specificity across all datasets. The authors claimed that the method requires fewer parameters and less FLOPs. However, in Table 2, there're other 3-4 methods are less complex and more lightweight than the proposed method.

Critical strengths

Consider spatial and frequency and design the two branches Possibly DCI before fusion is a good attempt.

Critical weaknesses

Incremental technical contribution, building a pipeline and not groundbreaking. Single domain, and within the domain, it seems the depth of the current issues in skin segmentation is unclear. The results are comparable with other methods, and the authors did not prove how the proposed method utilising boundary details / global structure. The rigour, robustness and generalisability of the proposed method is unclear.

Questions for authors

What is/are the best segmentation method(s) to date? Are they not considering boundary detail and global structure at design stage? What is the main novelty of your proposed method? What has been improved when compared to the state-of-the-art? How it performs on other image domain? Why is the specificity always lower than other methods?

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

Related papers

Discussion (0)

Sign in to comment.