SpectMamba: Integrating Frequency and State Space Models for Enhanced Medical Image Detection

Dong Yang; Liuzhi Yang; Wenjian Huang; Yao Wang; Zhen Qian; Zhi Qiao

REVIEW 5 major objections 5 minor 53 references

SpectMamba: Integrating Frequency and State Space Models for Enhanced Medical Image Detection

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SpectMamba claims that separating high- and low-frequency features and scanning along a Hilbert curve lets a state-space detector beat CNN, Transformer, and prior Mamba baselines on three medical benchmarks.

desk verdict A competent Mamba+frequency detector with a credible Brats gain, but the SOTA claim is undercut by an unfair DETR baseline and missing variance. read the letter →

arxiv 2509.01080 v1 pith:YKCRW2UB submitted 2025-09-01 cs.CV

Yao Wang , Dong Yang , Zhi Qiao , Wenjian Huang , Liuzhi Yang , Zhen Qian This is my paper

classification cs.CV

keywords medicalimagedetectionstate-spacemodelsMambafrequencybiasHilbertcurvescanninghybridspatial-frequencyattentionobjectX-rayandMRIabnormality

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

SpectMamba is presented as the first Mamba-based architecture built for medical image detection, and it is designed around a specific diagnosis: state-space models fail at this task for two fixable reasons. The first is frequency bias—low frequencies dominate learning, so fine edges and textures get lost—and the second is that flattening a 2D image into a 1D sequence destroys local spatial relationships. The paper answers with two components: a Hybrid Spatial-Frequency Attention (HSFA) block that learns high- and low-frequency features separately, and a Visual State-Space Module (VSSM) that scans the image along a Hilbert curve, which keeps neighboring pixels close in the sequence. On three public benchmarks—pneumonia X-rays, brain-tumor MRI slices, and pediatric wrist-fracture X-rays—SpectMamba reports higher mean average precision than the CNN, Transformer, and Mamba baselines it is compared against, while running at roughly twice VMamba's throughput. The result matters because, if the comparisons are fair, it suggests linear-complexity Mamba backbones can be both accurate and practical for high-resolution medical imaging.

What carries the argument

The central object is the Hybrid Spatial-Frequency Attention (HSFA) block paired with a Hilbert-scan Visual State-Space Module (VSSM). HSFA uses average pooling to split features into low- and high-frequency components, transforms them with a learnable DCT-domain matrix, and fuses them with multi-scale depthwise convolutions, so high-frequency edge and texture information is not erased by state-space frequency bias. VSSM then feeds the mixed features along a Hilbert curve—a locality-preserving space-filling curve—into selective state-space S6 blocks, so 1D sequence modeling retains 2D neighborhood structure. Together these two components carry the argument: one fixes what SSMs lose in fine d

What would settle it

Retrain the DETR-based baselines (DINO and DDQ) with their standard pretrained backbones and several random seeds on the same three datasets; if their mAP matches or passes SpectMamba's, the claimed SOTA advantage is an artifact of the no-pretraining protocol.

Watch

Extended reading notes

Core claim

SpectMamba's central discovery claim is that a state-space detector can reach or exceed CNN- and Transformer-based medical detectors if two structural problems are addressed explicitly. The HSFA block separates a feature map into low- and high-frequency branches using average pooling, applies a learnable DCT-domain transform to the frequency branches, and fuses them with multi-scale depthwise convolutions, so high-frequency lesion boundaries survive the frequency bias of SSMs. The VSSM then reads the fused features along a Hilbert curve—a space-filling curve that preserves 2D locality—before passing them through Mamba S6 blocks, so long-range context is modeled without destroying neighborhoo

Load-bearing premise

The main load-bearing assumption is that training every baseline from scratch for 200 epochs without pretrained weights is a fair and representative protocol; if the DETR-style baselines need pretraining to show their normal strength, the reported margins over them collapse.

Editorial extensions

If this is right

Mamba-style backbones can be used for medical detection, not just segmentation or classification, showing a path to linear-complexity global modeling in two-dimensional medical images.
Separating high- and low-frequency features before state-space scanning is a workable remedy for SSM frequency bias, preserving the edge and texture cues radiologists rely on.
Locality-preserving scans matter: a bidirectional Hilbert scan outperforms four-direction scans while using half the memory, suggesting other vision state-space models could drop the expensive four-direction habit.
The reported speed-accuracy combination—about double VMamba's throughput at similar or better mAP, and higher accuracy than ViT-L with 32% of its parameters—positions Mamba detectors as a practical option for high-resolution clinical images.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper does not test segmentation or landmark localization, but the same frequency-preservation and locality-preserving scan logic should transfer to any dense medical task where fine boundaries matter.
A multi-seed replication with error bars would be needed to confirm the small Graz margin of 0.48 mAP, which is not reported and could be within run-to-run noise.
A natural extension is to attach the HSFA block and Hilbert scan to other state-space vision backbones, independent of the FCOS detection head used here, to see whether the gains are general or specific to this detector.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Here's my take on SpectMamba. The paper is a reasonable engineering contribution: it builds a medical image detector on VMamba with two added pieces—a hybrid spatial-frequency attention block (HSFA) and a Hilbert-curve scanning scheme for the SSM—and it reports solid gains on three public datasets. The Brats result (75.34 mAP vs. VMamba's 71.64) is the strongest evidence that the architecture does something useful. The efficiency figure (twice VMamba's throughput) also looks real, though the FLOPs numbers in Table 2 make me want to see the measurement setup.

The soft spots are exactly where the reader's report puts them. The headline 'state-of-the-art' rests heavily on Table 3, where DDQ and DINO are trained from scratch for 200 epochs without pretrained backbones. That protocol is disclosed in Section 4.3, but it is not the recipe those models were designed for, and the paper's citation to [44] does not establish that from-scratch training is a fair or representative condition for them. The margins over DDQ/DINO are therefore not credible evidence of superiority. The comparison against VMamba, ResNet, and ViT in Table 1 is more fair—all appear to be trained under the same protocol—but even there the Graz gain is 0.48 mAP, which is within run-to-run noise given no seeds or error bars. The ablation table (Table 4) is also hard to read; the rows don't clearly indicate which components are on or off.

The paper is not a breakthrough, and its 'first Mamba-based medical detector' framing is a bit strong given Fusion-Mamba exists, but it is an honest engineering effort with plausible components and a clear experimental footprint on public data. The authors did not hide the baseline protocol; they argued for it. I think the central idea is sound enough to be worth a serious review, but the SOTA claim needs to be revised to a more modest claim about competitive performance under a controlled protocol, and the experiments need variance estimates and a fixed seed.

My recommendation: send it to peer review, not desk reject, but expect the reviewers to push for a fairer DETR comparison and for error bars. I would bring it to a reading group if someone wants to look at Mamba detectors for medical imaging; I'd cite it only if the code and the revised numbers hold up.

Referee Report

5 major / 5 minor

Summary. SpectMamba is a Mamba-based detection architecture for medical images. It combines a Hybrid Spatial-Frequency Attention (HSFA) block, which separates and recombines high- and low-frequency features via DCT/IDCT, with a Visual State-Space Module (VSSM) that uses Hilbert-curve scanning to preserve spatial locality. The detector is built on FPN and an FCOS-style head, and it is evaluated on RSNA pneumonia X-ray, BraTS brain MRI, and GRAZPEDWRI-DX wrist fracture X-ray datasets. The paper reports mAP gains over CNN, Transformer, and prior Mamba backbones, claims state-of-the-art performance against DDQ and DINO, and reports roughly 2x the throughput of VMamba.

Significance. If the experimental claims were robust, SpectMamba would make a useful contribution: it addresses two known weaknesses of Mamba vision models—loss of local spatial structure in 1D scanning and spectral bias toward low frequencies—with simple, feasible mechanisms (Hilbert scanning and high/low-frequency separation). The efficiency result (275 vs. 136 throughput at comparable parameter count) is also interesting. However, the current evidence does not establish the headline SOTA claim. The comparison with DDQ/DINO uses a from-scratch, no-pretraining protocol that is not representative of those methods, and the gains over the strongest Mamba baseline on Graz (0.48 mAP) and PenD (1.07 mAP) are small enough to lie within run-to-run variation, especially since no seeds or error bars are reported. The architectural ideas are worth pursuing, but the evaluation and claims need substantial revision.

major comments (5)

[Section 4.3, Table 3] The SOTA comparison is unfairly configured. DDQ and DINO are trained from scratch with no pretrained weights for only 200 epochs, whereas these DETR variants are designed and normally evaluated with ImageNet-pretrained backbones and substantially longer schedules (e.g., DINO typically uses 500 epochs). The reported mAP values for DDQ/DINO (7.47/7.55 on PenD, 30.22/40.41 on Brats, 41.13/42.55 on Graz) are far below what these architectures achieve under their standard recipe. The citation [44] does not justify removing pretraining; it is about specialized DETR components, not about abandoning pretrained backbones. The claim "SpectMamba consistently outperforms these SOTA models" is therefore not supported. Please either compare with DDQ/DINO using their standard pretraining and training schedules, or explicitly label these as "from-scratch, restricted-protocol" comparisons and remove the
[Section 4.1, Tables 1 and 4-5] No seeds, error bars, or significance tests are reported anywhere. The Graz improvement over VMamba is 0.48 mAP (77.89 vs. 77.41), and the PenD improvement is 1.07 mAP (25.26 vs. 24.19); both are plausibly within run-to-run noise for a 200-epoch training without a fixed seed. The ablation tables (Tables 4 and 5) also lack variance information. Please report mean ± std over at least three random seeds and state whether the observed differences are statistically significant. At minimum, a fixed seed should be specified so that results are reproducible.
[Table 4] The ablation table is ambiguous. The column headers are "Bidi-scan," "Hilbert curve," "LH-info," and "spatial," but the rows use check marks without an explicit mapping. For instance, the first row has a check under "Bidi-scan" only; the third row has checks under "Hilbert curve" and "LH-info"; it is not clear what configuration each row represents or how the "LH-info" and "spatial" components differ from the HSFA block described in Section 3.2. Please replace the cryptic check-mark table with an explicit list of configurations, e.g., "Baseline (Bidi-scan only)," "Bidi-scan + LH-info," etc. Also clarify the relationship between Table 5's "Hilbert-BiDir" and the bidirectional baseline in Table 4.
[Section 3.2, Eq. (5)] The frequency-domain mechanism is not sufficiently specified for reproduction. Equation (5) writes Q_h^f, Q_l^f = IDCT(E_f(DCT(F_h^f, F_l^f))), but it is unclear whether DCT is applied jointly to a concatenation or separately to high and low components, what the dimensions of the DCT and IDCT are, and what shape the learnable matrix E_f has (per-channel? per spatial block?). The text also inconsistently refers to the "FH-info module" in Section 3.2 and the "LH-Info Separator" in Figure 1. Please provide exact tensor shapes and processing steps for the DCT-based branch and for the high/low decomposition.
[Table 2, Section 4.2] The efficiency claim needs clarification. Table 2 is captioned "Train Throughput," and the text says SpectMamba "operating at twice the speed of VMamba" (275 vs. 136). Please state explicitly whether these are training or inference throughput, and under what hardware/batch conditions they were measured. Also, SpectMamba's FLOPs (77.5 G) are roughly double VMamba's (38.0 G) even though its parameter count is slightly lower; the paper should interpret "efficiency" in terms of both throughput and FLOPs, not only throughput.

minor comments (5)

[Section 4.1, Baselines] PCViT is listed as a Transformer-based baseline, but no PCViT result appears in Table 1. Please either add the result or remove the method from the list.
[Section 3.1 and 3.2] Typos: "the the convolution operation" should be "the convolution operation," and "bilnear interpolation" should be "bilinear interpolation." Also, in Section 3.2 the component is called "FH-info module" in one place and "LH-Info Separator" in Figure 1; unify the terminology.
[References] In the Introduction, "FreqMamba[5,53]" cites two references; reference [5] is not FreqMamba but a remote-sensing segmentation paper. Please correct the citation. Also, Section 3.4 cites [40] for λ1=λ2=1, but reference [40] is a LiDAR object-detection paper; the appropriate citation is likely the FCOS paper [41].
[Section 4.5, Figure 2] The text says the ERFs are visualized "prior to training," but the caption and surrounding explanation do not say what is being compared or why untrained ERFs are informative. Please clarify.
[Section 4.4] The sentence "Prior work[25] has shown that bidirectional scanning is more computationally efficient and yields better classification accuracy than cascade scanning" is not directly about the ablation being performed; please connect this observation to the current results or remove it.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical architecture paper with held-out evaluations; no prediction reduces to a fit or self-citation.

full rationale

SpectMamba is an empirical architecture paper. Its central claims are that the proposed HSFA block, VSSM, and Hilbert-curve scanning improve medical image detection, supported by held-out test results on three public benchmarks. There is no analytic derivation in which a predicted quantity is defined in terms of the fitted input: the architecture components are described by their forward computations (e.g., Eqs. 4-8) and the loss is a standard FCOS loss (Eq. 9), not a quantity fitted to the test set. The comparisons to DDQ and DINO involve a training-protocol choice (no pretraining, 200 epochs) that may be questionable as a fairness matter, but this is not circular reasoning; it does not make the reported SpectMamba results equivalent to the experimental inputs. I also checked for load-bearing self-citations or imported uniqueness theorems. The paper cites prior work for Mamba, frequency bias, and scanning ideas, but none of these citations are by the present authors, and none are used to forbid alternatives or to define the proposed method into existence. The ablation study compares variants on the validation set, which is a standard empirical practice, not a circular reduction. Overall, no circular step can be identified by quoting specific equations or fitted-parameter renaming.

Assumptions & free parameters 4 free parameters · 4 assumptions · 3 invented entities

The paper's central claim relies on standard deep learning assumptions (SSM effectiveness, frequency bias, Hilbert locality, FCOS head) and on several hand-chosen hyperparameters. No analytic derivation is claimed, so the ledger is mostly checking the architectural premises.

free parameters (4)

Kernel size set K={3,5} in HSFA
Chosen by hand for multi-scale spatial extraction; no ablation over kernel sizes is performed (Section 3.2, Eq. 4).
LH-info downsample rate = 2
Downsampling rate and step size set to 2, upsampling via bilinear interpolation, following OctConv-style design without ablation (Section 3.2).
Loss weights lambda1, lambda2 = 1
Set to 1 following FCOS [40]; no tuning reported (Section 3.4).
Network stages and FPN levels = 4 stages, 5 FPN levels
Architecture choices inherited from common detection backbones; not justified by experiments (Section 3.2, 3.4).

assumptions (4)

domain assumption Mamba-style SSMs can model long-range dependencies in images with linear complexity.
The whole VSSM relies on this property (Section 3.1, 3.3).
domain assumption Frequency bias in SSMs suppresses high-frequency details and harms medical detection.
Motivates the HSFA block; cited from [46,34] but not directly measured in this paper (Section 1, 3.2).
domain assumption Hilbert curve scan preserves local spatial dependencies better than bidirectional or cascade scans.
The core of the proposed scanning method; only qualitatively supported by ERF visualizations (Section 4.5, Fig 2).
domain assumption FCOS detection loss and FPN are suitable for medical object detection.
Used as the detection head; standard practice in the field (Section 3.4).

invented entities (3)

Hybrid Spatial-Frequency Attention (HSFA) block
purpose: Extract and combine spatial and high/low frequency features to reduce frequency bias
Only evaluated within this paper's experiments; no external validation.
Hilbert-SS2D scanning module
purpose: Perform Mamba scanning along Hilbert curve to preserve locality
Performance evidence is internal; no external benchmark or analysis.
Low-High Frequency Domain Information Separator (LH-Info Separator)
purpose: Decompose feature maps into low and high frequency components via DCT
A submodule of HSFA; not independently evaluated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SpectMamba: Integrating Frequency and State Space Models for Enhanced Medical Image Detection." pith.science (2026). https://pith.science/paper/YKCRW2UB

@misc{pith2026250901080,
  author       = {Pith},
  title        = {Pith review of: SpectMamba: Integrating Frequency and State Space Models for Enhanced Medical Image Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YKCRW2UB}},
  note         = {Machine review of arXiv:2509.01080}
}

read the original abstract

Abnormality detection in medical imaging is a critical task requiring both high efficiency and accuracy to support effective diagnosis. While convolutional neural networks (CNNs) and Transformer-based models are widely used, both face intrinsic challenges: CNNs have limited receptive fields, restricting their ability to capture broad contextual information, and Transformers encounter prohibitive computational costs when processing high-resolution medical images. Mamba, a recent innovation in natural language processing, has gained attention for its ability to process long sequences with linear complexity, offering a promising alternative. Building on this foundation, we present SpectMamba, the first Mamba-based architecture designed for medical image detection. A key component of SpectMamba is the Hybrid Spatial-Frequency Attention (HSFA) block, which separately learns high- and low-frequency features. This approach effectively mitigates the loss of high-frequency information caused by frequency bias and correlates frequency-domain features with spatial features, thereby enhancing the model's ability to capture global context. To further improve long-range dependencies, we propose the Visual State-Space Module (VSSM) and introduce a novel Hilbert Curve Scanning technique to strengthen spatial correlations and local dependencies, further optimizing the Mamba framework. Comprehensive experiments show that SpectMamba achieves state-of-the-art performance while being both effective and efficient across various medical image detection tasks.

Figures

Figures reproduced from arXiv: 2509.01080 by the authors.

**Figure 3.** Visualized examples of SpectMamba and comparative methods, where red boxes indicate ground truths and blue boxes represent predictions [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 36 canonical work pages

[44]

Understanding differences in applying DETR to natural and medical images

Xu, Y., Shen, Y., Fernandez-Granda, C., Heacock, L., Geras, K.J.: Understand- ing differences in applying DETR to natural and medical images. arXiv preprint arXiv:2405.17677 (2024) 16 Authors Suppressed Due to Excessive Length

work page Pith review arXiv 2024
[1]

Scientific data 4(1), 1–13 (2017)

Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J.S., Freymann, J.B., Farahani, K., Davatzikos, C.: Advancing the cancer genome atlas glioma MRI collections with expert segmentation labels and radiomic features. Scientific data 4(1), 1–13 (2017)

work page 2017
[2]

arXiv preprint arXiv:1811.02629 (2018)

Bakas, S., Reyes, M., Jakab, A., Bauer, S., Rempfler, M., Crimi, A., Shinohara, R.T., Berger, C., Ha, S.M., Rozycki, M., et al.: Identifying the best machine learn- ing algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the brats challenge. arXiv preprint arXiv:1811.02629 (2018)

arXiv 2018
[3]

Prentice-Hall, Inc

Brigham, E.O.: The fast Fourier transform and its applications. Prentice-Hall, Inc. (1988)

work page 1988
[4]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Cai, Z., Vasconcelos, N.: Cascade R-CNN: Delving into high quality object de- tection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6154–6162 (2018)

work page 2018
[5]

Remote Sensing Image Segmentation Using Vision Mamba and Multi-Scale Multi-Frequency Feature Fusion

Cao, Y., Liu, C., Wu, Z., Yao, W., Xiong, L., Chen, J., Huang, Z.: Remote sensing image segmentation using vision mamba and multi-scale multi-frequency feature fusion. arXiv preprint arXiv:2410.05624 (2024)

work page Pith review arXiv 2024
[6]

IEEE Transactions on Image Processing31, 4609– 4621 (2022)

Chen, J., Yu, L., Wang, W.: Hilbert space filling curve based scan-order for point cloud attribute compression. IEEE Transactions on Image Processing31, 4609– 4621 (2022)

work page 2022
[7]

In: Proceedings of the IEEE/CVF international confer- ence on computer vision

Chen, Y., Fan, H., Xu, B., Yan, Z., Kalantidis, Y., Rohrbach, M., Yan, S., Feng, J.: Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution. In: Proceedings of the IEEE/CVF international confer- ence on computer vision. pp. 3435–3444 (2019)

work page 2019

Show all 53 references

[8]

arXiv preprint arXiv:2405.21060 (2024)

Dao, T., Gu, A.: Transformers are SSMs: Generalized models and efficient algo- rithms through structured state space duality. arXiv preprint arXiv:2405.21060 (2024)

2024 arXiv
[9]

multimedia Tools and Applications 82(6), 9243–9275 (2023)

Diwan, T., Anirudh, G., Tembhurne, J.V.: Object detection using YOLO: Chal- lenges, architectural successors, datasets and applications. multimedia Tools and Applications 82(6), 9243–9275 (2023)

2023
[10]

arXiv preprint arXiv:2404.09146 (2024)

Dong, W., Zhu, H., Lin, S., Luo, X., Shen, Y., Liu, X., Zhang, J., Guo, G., Zhang, B.: Fusion-mamba for cross-modality object detection. arXiv preprint arXiv:2404.09146 (2024)

2024 arXiv
[11]

arXiv preprint arXiv:2010.11929 (2020)

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is 14 Authors Suppressed Due to Excessive Length worth 16x16 words: Transformers for image recognition at scale. arXiv p...

2010 arXiv
[12]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops

Gabruseva, T., Poplavskiy, D., Kalinin, A.: Deep learning for automatic pneumonia detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 350–351 (2020)

2020
[13]

Galdran, A., Carneiro, G., González Ballester, M.A.: Balanced-mixup for highly imbalanced medical image classification. In: Medical Image Computing and Com- puter Assisted Intervention–MICCAI 2021: 24th International Conference, Stras- bourg, France, September 27–October 1, 20...

2021
[14]

arXiv preprint arXiv:2312.00752 (2023)

Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023)

2023 arXiv
[15]

Gu, A., Dao, T., Ermon, S., Rudra, A., Ré, C.: Hippo: Recurrent memory with optimalpolynomialprojections.Advancesinneuralinformationprocessingsystems 33, 1474–1487 (2020)

2020
[16]

arXiv preprint arXiv:2111.00396 (2021)

Gu, A., Goel, K., Ré, C.: Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396 (2021)

2021 arXiv
[17]

Advances in neural information processing systems34, 572–585 (2021)

Gu, A., Johnson, I., Goel, K., Saab, K., Dao, T., Rudra, A., Ré, C.: Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems34, 572–585 (2021)

2021
[18]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

He, C., Li, R., Li, S., Zhang, L.: Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8417–8427 (2022)

2022
[19]

In: European Conference on Computer Vision

Hu, V.T., Baumann, S.A., Gui, M., Grebenkova, O., Ma, P., Fischer, J., Ommer, B.: Zigma: A dit-style zigzag mamba diffusion model. In: European Conference on Computer Vision. pp. 148–166. Springer (2025)

2025
[20]

arXiv preprint arXiv:2305.12073 (2023)

Lee, M.: Gelu activation function in deep learning: a comprehensive mathematical analysis and performance. arXiv preprint arXiv:2305.12073 (2023)

2023 arXiv
[21]

IEEE Transactions on Geoscience and Remote Sensing 62, 1–15 (2024)

Li, J., Tian, P., Song, R., Xu, H., Li, Y., Du, Q.: Pcvit: A pyramid convo- lutional vision transformer detector for object detection in remote-sensing im- agery. IEEE Transactions on Geoscience and Remote Sensing 62, 1–15 (2024). https://doi.org/10.1109/TGRS.2024.3360456

2024
[22]

In: European conference on computer vision

Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: European conference on computer vision. pp. 280–296. Springer (2022)

2022
[23]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Lin, T.Y., Dollár, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117–2125 (2017)

2017
[24]

Medical image analysis42, 60–88 (2017)

Litjens, G., Kooi, T., Bejnordi, B.E., Setio, A.A.A., Ciompi, F., Ghafoorian, M., Van Der Laak, J.A., Van Ginneken, B., Snchez, C.I.: A survey on deep learning in medical image analysis. Medical image analysis42, 60–88 (2017)

2017
[25]

Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model (2024), https://arxiv.org/abs/2401.10166

2024 arXiv
[26]

Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer:Hierarchical visiontransformerusingshiftedwindows.In:Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012–10022 (2021)

2021
[27]

Advances in neural information processing systems 29 (2016) Title Suppressed Due to Excessive Length 15

Luo, W., Li, Y., Urtasun, R., Zemel, R.: Understanding the effective receptive field in deep convolutional neural networks. Advances in neural information processing systems 29 (2016) Title Suppressed Due to Excessive Length 15

2016
[28]

arXiv preprint arXiv:2206.13947 (2022)

Mehta, H., Gupta, A., Cutkosky, A., Neyshabur, B.: Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947 (2022)

2022 arXiv
[29]

IEEE transactions on medical imaging 34(10), 1993–2024 (2014)

Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., Burren, Y., Porz, N., Slotboom, J., Wiest, R., et al.: The multimodal brain tumor image segmentation benchmark (BRATS). IEEE transactions on medical imaging 34(10), 1993–2024 (2014)

1993
[30]

GeoInformatica 7, 179–209 (2003)

Mokbel, M.F., Aref, W.G., Kamel, I.: Analysis of multi-dimensional space-filling curves. GeoInformatica 7, 179–209 (2003)

2003
[31]

Scientific data9(1), 222 (2022)

Nagy, E., Janisch, M., Hržić, F., Sorantin, E., Tschauner, S.: A pediatric wrist trauma X-ray dataset (grazpedwri-dx) for machine learning. Scientific data9(1), 222 (2022)

2022
[32]

In: Proceedings of the 1986 ACM SIGMOD international conference on Management of data

Orenstein, J.A.: Spatial query processing in an object-oriented database system. In: Proceedings of the 1986 ACM SIGMOD international conference on Management of data. pp. 326–336 (1986)

1986
[33]

Advances in Neural Information Processing Systems35, 14541–14554 (2022)

Pan, Z., Cai, J., Zhuang, B.: Fast vision transformers with hilo attention. Advances in Neural Information Processing Systems35, 14541–14554 (2022)

2022
[34]

In: International conference on machine learning

Rahaman, N., Baratin, A., Arpit, D., Draxler, F., Lin, M., Hamprecht, F., Ben- gio, Y., Courville, A.: On the spectral bias of neural networks. In: International conference on machine learning. pp. 5301–5310. PMLR (2019)

2019
[35]

Advances in neural information processing systems34, 980–993 (2021)

Rao, Y., Zhao, W., Zhu, Z., Lu, J., Zhou, J.: Global filter networks for image clas- sification. Advances in neural information processing systems34, 980–993 (2021)

2021
[36]

arXiv preprint arXiv:2402.02491 (2024)

Ruan, J., Xiang, S.: Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491 (2024)

2024 arXiv
[37]

In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

Shi, D.: Transnext: Robust foveal visual perception for vision transformers. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 17773–17783 (2024)

2024
[38]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: Rsn: Range sparse net for efficient, accurate lidar 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5725–5734 (2021)

2021
[39]

In: International Conference on Med- ical Image Computing and Computer-Assisted Intervention

Tang, Z., Jiang, C., Cui, Z., Shen, D.: Hf-resdiff: High-frequency-guided residual diffusion for multi-dose PET reconstruction. In: International Conference on Med- ical Image Computing and Computer-Assisted Intervention. pp. 372–381. Springer (2024)

2024
[40]

Advances in Neural Information Processing Systems 35, 34899–34911 (2022)

Tian, Z., Chu, X., Wang, X., Wei, X., Shen, C.: Fully convolutional one-stage 3D object detection on lidar range images. Advances in Neural Information Processing Systems 35, 34899–34911 (2022)

2022
[41]

IEEE transactions on pattern analysis and machine intelligence44(4), 1922–1933 (2020)

Tian, Z., Shen, C., Chen, H., He, T.: Fcos: A simple and strong anchor-free object detector. IEEE transactions on pattern analysis and machine intelligence44(4), 1922–1933 (2020)

1922
[42]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Wang, J., Zhu, W., Wang, P., Yu, X., Liu, L., Omar, M., Hamid, R.: Selective structured state-spaces for long-form video understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6387– 6397 (2023)

2023
[43]

Computers in Biology and Medicine170, 107983 (2024)

Wei, K., Kong, W., Liu, L., Wang, J., Li, B., Zhao, B., Li, Z., Zhu, J., Yu, G.: CT synthesis from MR images using frequency attention conditional generative adversarial network. Computers in Biology and Medicine170, 107983 (2024)

2024
[45]

Computers in Biology and Medicine163, 107149 (2023)

Xu, Z., Zhang, X., Zhang, H., Liu, Y., Zhan, Y., Lukasiewicz, T.: Efpn: Effective medical image detection using feature pyramid fusion enhancement. Computers in Biology and Medicine163, 107149 (2023)

2023
[46]

arXiv preprint arXiv:2410.02035 (2024)

Yu, A., Lyu, D., Lim, S.H., Mahoney, M.W., Erichson, N.B.: Tuning frequency bias of state space models. arXiv preprint arXiv:2410.02035 (2024)

2024 arXiv
[47]

In: Proceedings of the 24th ACM international conference on Multimedia

Yu, J., Jiang, Y., Wang, Z., Cao, Z., Huang, T.: Unitbox: An advanced object detection network. In: Proceedings of the 24th ACM international conference on Multimedia. pp. 516–520 (2016)

2016
[48]

arXiv preprint arXiv:2203.03605 (2022)

Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino:DETRwithimproveddenoisinganchorboxesforend-to-endobjectdetection. arXiv preprint arXiv:2203.03605 (2022)

2022 arXiv
[49]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Zhang, S., Wang, X., Wang, J., Pang, J., Lyu, C., Zhang, W., Luo, P., Chen, K.: Dense distinct query for end-to-end object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7329–7338 (2023)

2023
[50]

Elec- tronics 9(3), 537 (2020)

Zhao, L., Li, S.: Object detection algorithm based on improved YOLOv3. Elec- tronics 9(3), 537 (2020)

2020
[51]

arXiv preprint arXiv:2401.09417 (2024)

Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024)

2024 arXiv
[52]

Heliyon 10(19) (2024)

Zhu, Q., Cai, Y., Fang, Y., Yang, Y., Chen, C., Fan, L., Nguyen, A.: Samba: Semantic segmentation of remotely sensed images with state space model. Heliyon 10(19) (2024)

2024
[53]

In: Proceedings of the 32nd ACM International Conference on Multimedia

Zou,Z., Yu, H.,Huang, J., Zhao, F.: Freqmamba: Viewingmamba froma frequency perspective for image deraining. In: Proceedings of the 32nd ACM International Conference on Multimedia. pp. 1905–1914 (2024)

1905

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [44]

Understanding differences in applying DETR to natural and medical images

Xu, Y., Shen, Y., Fernandez-Granda, C., Heacock, L., Geras, K.J.: Understand- ing differences in applying DETR to natural and medical images. arXiv preprint arXiv:2405.17677 (2024) 16 Authors Suppressed Due to Excessive Length

work page Pith review arXiv 2024

[2] [1]

Scientific data 4(1), 1–13 (2017)

Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J.S., Freymann, J.B., Farahani, K., Davatzikos, C.: Advancing the cancer genome atlas glioma MRI collections with expert segmentation labels and radiomic features. Scientific data 4(1), 1–13 (2017)

work page 2017

[3] [2]

arXiv preprint arXiv:1811.02629 (2018)

Bakas, S., Reyes, M., Jakab, A., Bauer, S., Rempfler, M., Crimi, A., Shinohara, R.T., Berger, C., Ha, S.M., Rozycki, M., et al.: Identifying the best machine learn- ing algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the brats challenge. arXiv preprint arXiv:1811.02629 (2018)

arXiv 2018

[4] [3]

Prentice-Hall, Inc

Brigham, E.O.: The fast Fourier transform and its applications. Prentice-Hall, Inc. (1988)

work page 1988

[5] [4]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Cai, Z., Vasconcelos, N.: Cascade R-CNN: Delving into high quality object de- tection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6154–6162 (2018)

work page 2018

[6] [5]

Remote Sensing Image Segmentation Using Vision Mamba and Multi-Scale Multi-Frequency Feature Fusion

Cao, Y., Liu, C., Wu, Z., Yao, W., Xiong, L., Chen, J., Huang, Z.: Remote sensing image segmentation using vision mamba and multi-scale multi-frequency feature fusion. arXiv preprint arXiv:2410.05624 (2024)

work page Pith review arXiv 2024

[7] [6]

IEEE Transactions on Image Processing31, 4609– 4621 (2022)

Chen, J., Yu, L., Wang, W.: Hilbert space filling curve based scan-order for point cloud attribute compression. IEEE Transactions on Image Processing31, 4609– 4621 (2022)

work page 2022

[8] [7]

In: Proceedings of the IEEE/CVF international confer- ence on computer vision

Chen, Y., Fan, H., Xu, B., Yan, Z., Kalantidis, Y., Rohrbach, M., Yan, S., Feng, J.: Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution. In: Proceedings of the IEEE/CVF international confer- ence on computer vision. pp. 3435–3444 (2019)

work page 2019