WING: A Window-Prior-Based Generative Network with Gated Inception for Cross-Modality CT Synthesis

Andreas Maier; Chengze Ye; Daiqi Liu; Fuxin Fan; Siming Bayer; Siyuan Mei; Yan Xia; Yipeng Sun; Yixing Huang; Zirong Li

REVIEW 3 major objections 6 minor 30 references

WING: A Window-Prior-Based Generative Network with Gated Inception for Cross-Modality CT Synthesis

T0 review · 3 major / 6 minor · reviewed 2026-07-08 · glm-5.2

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

Pith's one-line read Split CT Into Three Windows, Fuse With a Transformer, Beat SoTA

desk verdict Window-decomposition for CT synthesis is a clean idea, but the ablation confounds windowing with added Transformer capacity, leaving the central thesis under-tested. read the letter →

arxiv 2607.06234 v1 pith:ZDAQLNGN submitted 2026-07-07 cs.CV

Siyuan Mei , Yan Xia , Yipeng Sun , Siming Bayer , Zirong Li , Chengze Ye , Daiqi Liu , Fuxin Fan

show 2 more authors

Yixing Huang Andreas Maier

This is my paper · ORCID

classification cs.CV PACS 87.57.-s87.57.N87.57.C

keywords windowedwingcross-modalitydistributionsgatedgenerativeinceptionintensities

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

This paper argues that the standard approach to synthetic CT generation—regressing the full Hounsfield unit range in one shot—is fundamentally flawed because CT intensity distributions are long-tailed: abundant soft-tissue voxels dominate the loss, and sparse but clinically critical structures like lung and bone get averaged away. The authors propose that instead of treating CT synthesis as a single-range regression problem, it should be decomposed into three non-overlapping radiological windows (lung, soft tissue, bone), each of which has a smoother, more learnable distribution. They build a generator that predicts these three windows separately, then fuse them back into a full-range CT using a differentiable soft fusion operator whose blending masks are derived from the soft-tissue window prediction, followed by a Transformer that learns a residual correction on top of the coarse fusion. The central claim is that this window-decomposed reformulation is superior to direct full-range regression, and they support it with state-of-the-art results on both MRI-to-CT and CBCT-to-CT benchmarks using the SynthRAD2025 dataset, outperforming the previous best by 1.58 and 1.99 MAE respectively, while using only 19.29M parameters and handling multiple anatomies with a single model.

What carries the argument

The central mechanism is the window-decompose-then-fuse pipeline: CT volumes are split into three fixed radiological windows (lung: WL -600, WW 800; soft tissue: WL 0, WW 400; bone: WL 1100, WW 1800), predicted as separate channels by a Gated Inception Generator using anisotropic depthwise kernels (3x3x3, 3x9x1, 3x1x9) with gated branch fusion, then recombined by a differentiable soft fusion function (Eq. 3-4) whose spatial masks are derived from the soft-tissue window prediction, with a Transformer learning residual corrections on the coarse fused output.

What would settle it

If the soft-tissue window prediction contains systematic errors near its intensity boundaries (e.g., at the lung-soft-tissue or soft-tissue-bone transition), the derived fusion masks for lung and bone regions would be spatially misplaced, producing structured artifacts in the fused CT that are worse than those produced by direct full-range regression. A targeted perturbation of soft-tissue window boundary predictions should degrade final CT quality disproportionately compared to perturbations in the lung or bone windows themselves.

Watch

Extended reading notes

Core claim

The key finding is that decomposing CT regression into clinically motivated, non-overlapping intensity windows—lung, soft tissue, and bone—transforms an intractable long-tailed regression problem into three tractable ones with smoother distributions, and that a differentiable soft fusion operator derived from the soft-tissue window can reconstruct these windows back into a faithful full-range CT. The per-window ablation shows the largest gains precisely in the long-tailed extremes (lung and bone), confirming that the improvement comes from rebalancing the learning signal rather than from any single architectural trick. The Gated Inception Block and Fuse-and-Refine Transformer each contribute

Load-bearing premise

The soft fusion operator derives all fusion masks solely from the soft-tissue window prediction, assuming that the soft-tissue window's intensity bounds reliably characterize the transitions between all three windows. If the predicted soft-tissue window has errors at its boundaries, the fusion masks for lung and bone regions will be incorrect, propagating artifacts into the final CT. No sensitivity analysis to soft-tissue prediction errors at window boundaries is provided.

Editorial extensions

If this is right

If window decomposition generalizes, the same principle could apply to any regression target with long-tailed, multi-modal distributions—such as depth maps with mixed indoor/outdoor ranges or spectral imaging with distinct material bands.
The soft fusion operator's reliance on the soft-tissue window as the sole source of fusion masks creates a single point of failure: systematic errors in soft-tissue prediction at intensity boundaries would propagate artifacts into lung and bone regions of the final CT.
The fixed window settings follow standard radiological practice but may not be optimal for non-standard acquisition protocols or unusual anatomy, suggesting that learned or adaptive window boundaries could yield further gains.
The approach eliminates the need for external segmentation models or geometric guidance labels, which could simplify clinical deployment of synthetic CT in adaptive radiotherapy workflows where annotated data is scarce.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The main thing to know: this paper reformulates sCT generation as regression on three radiological windows (lung, soft tissue, bone) with a differentiable soft-fusion operator to reconstruct full-range CT. That reformulation is genuinely new for this subfield and well-motivated — the long-tailed HU distribution is a real problem, and decomposing it into smoother sub-distributions is a natural fix. The gains over MedNeXt on SynthRAD2025 are consistent across both tasks and all four metrics, with the largest improvements in lung and bone MAE (Table 1b), which is exactly where you'd expect windowing to help. The model is compact at 19.29M parameters. The per-window ablation in Table 1(b) is the strongest evidence that windowing itself does useful work — WING beats MedNeXt in every window, not just on aggregate. Credit where it's due: the idea is clean, the experimental scale is solid, and the results are not cherry-picked. The soft-fusion operator (Eqs. 3–4) is parameter-free aside from τ, which is a single scalar set to 0.2. That's good engineering. Now the soft spots. The stress-test concern about Table 2 is valid and is the most important issue. The +FRT row introduces three changes at once: windowed regression targets, the soft-fusion operator, and a Transformer refinement module with extra parameters. The paper attributes the improvement to 'window fusion and refinement strategies,' but you can't separate the window prior from the added architectural capacity. A proper ablation would apply the same Transformer residual to direct full-range regression. Without that, the claim that window decomposition is superior to direct regression rests on Table 1(b) alone, which is suggestive but not conclusive since the model was also trained with window-specific loss terms on a more powerful architecture. The reader's concern about fusion masks deriving solely from the soft-tissue window is structurally fair but probably minor in practice — the soft-tissue window boundaries are where the transitions actually occur, and τ=0.2 gives reasonable blending width. No sensitivity analysis is provided, but this is a secondary issue. More practically: no statistical significance tests despite standard deviations (±3.5–4.5 HU on MAE) that are large relative to the improvements (~1.5–2.0 HU), and no code or weights released. No downstream dose calculation, which is the clinical endpoint. These are addressable. This paper is for researchers working on medical image synthesis who care about incorporating domain priors into architecture design. The window-decomposition idea will likely be adopted regardless of whether the FRT module is the real driver. It deserves a serious referee who should ask for the missing ablation (Transformer on direct regression) and significance tests. If the authors can show windowing helps even without the Transformer, the contribution is solid.

Referee Report

3 major / 6 minor

Summary. This paper proposes WING, a window-prior-based generative network for cross-modality CT synthesis (MRI-to-CT and CBCT-to-CT). The core idea is to decompose the full-range CT regression target into three non-overlapping radiological windows (lung, soft tissue, bone) to mitigate the long-tailed distribution of Hounsfield units. The architecture uses a Gated Inception Generator (GIG) to produce multi-window predictions, a Fuse-and-Refine Transformer (FRT) to aggregate the windowed outputs via a differentiable soft fusion function and learn residuals for refinement, and a joint adversarial objective. Evaluated on the SynthRAD2025 dataset, the method achieves state-of-the-art performance on both tasks while using only 19.29M parameters and supporting multi-anatomy synthesis with a single model.

Significance. The paper presents a well-motivated approach to CT synthesis that leverages domain-specific priors (radiological windowing) rather than relying solely on architectural scaling. The use of standard radiological window settings (WL/WW) from external clinical references, rather than data-fitted parameters, is a strength that reduces circularity concerns. The method achieves strong empirical results on a large-scale public benchmark (SynthRAD2025) with standard metrics (MAE, MS-SSIM, PSNR, DICE), and the parameter efficiency (19.29M) is notable. The per-window MAE analysis in Table 1(b) provides useful insight into where the gains occur. The soft fusion operator (Eq. 3-4) is a clean, parameter-free mechanism for reconstructing the full-range CT from windowed predictions.

major comments (3)

Table 2, +FRT row: The ablation confounds windowed regression with the Transformer refinement module. The +FRT step introduces three changes simultaneously: (1) switching from direct full-range regression to three-channel windowed regression targets, (2) the soft fusion operator (Eq. 3-4), and (3) a Transformer-based residual refinement module with additional parameters and self-attention. The improvement attributed to FRT (0.79 MAE MRI, 1.50 MAE CBCT) could be driven by the added Transformer capacity rather than by the window decomposition itself. The paper's central thesis is that 'decomposing CT regression into windowed targets with structured fusion is superior to direct full-range regression,' but this claim is not isolated by the ablation. An additional ablation applying the Transformer residual refinement to direct full-range regression (same module, no windowing) would be needed.
Section 2.1, Eq. (3): The soft fusion operator derives fusion masks solely from the soft-tissue window prediction y(soft), assuming that the soft-tissue window's intensity bounds reliably characterize transitions between all three windows including lung and bone. If the predicted soft-tissue window has errors at its boundaries, the fusion masks for lung and bone regions will be incorrect, propagating artifacts into the final fused CT. No sensitivity analysis to soft-tissue prediction errors at window boundaries is provided. This is load-bearing because the entire reconstruction pipeline depends on the fidelity of this single-channel mask derivation.
Table 1(a): No statistical significance tests are reported for the metric improvements over baselines. The improvements over MedNeXt (1.58 MAE MRI-to-CT, 1.99 MAE CBCT-to-CT) are meaningful in absolute terms, but the standard deviations are large (e.g., ±20.77 for WING vs. ±23.53 for MedNeXt on MRI-to-CT MAE). Without paired statistical tests, it is unclear whether these improvements are statistically significant.

minor comments (6)

Section 2.1, Eq. (3): The notation uses y(soft) but the stacked representation in Eq. (2) uses y(soft) with a superscript. Consistent notation (superscript vs. parenthetical) would improve readability.
Section 2.2, GIG description: The text mentions 'input channels are evenly split into four branches' but does not specify how the identity branch is handled when channel counts are not divisible by four. Clarification would help reproducibility.
Table 1(b): The per-window MAE values for WING average (75.45 MRI, 49.94 CBCT) do not match the system-level MAE in Table 1(a) (61.92 MRI, 48.58 CBCT). This is presumably because Table 1(b) computes MAE only over voxels within clipped intensity ranges while Table 1(a) computes over all voxels, but this distinction should be stated explicitly.
Section 3.2: The loss weights are given as a ratio (λ_win : λ_ct : λ_lpips : λ_adv = 1 : 3 : 0.4 : 0.1) but the absolute values are not specified. Providing absolute values would aid reproducibility.
Figure 2: The 'Soft fuse' block is referenced but the connection between the soft fusion path and the Transformer path in FRT could be clearer. A reader may need to cross-reference Eq. (4) multiple times to understand the data flow.
Section 3.3: The DICE metric is computed with a 'fixed pre-trained segmentation model' but details about this model (architecture, training data, potential bias) are not provided. A brief reference or description would strengthen the evaluation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful and constructive review. The referee raises three substantive points: (1) the FRT ablation confounds windowed regression with Transformer-based refinement, (2) the soft fusion operator's dependence on the soft-tissue window prediction lacks sensitivity analysis, and (3) no statistical significance tests are reported for the system-level comparisons. We agree that all three points identify genuine gaps in the current manuscript and will address each in the revision. Specifically, we will conduct an additional ablation isolating the windowed regression contribution from the Transformer refinement, provide a sensitivity analysis for the soft fusion operator, and report paired statistical significance tests for all baseline comparisons.

read point-by-point responses

Referee: Table 2, +FRT row: The ablation confounds windowed regression with the Transformer refinement module. The +FRT step introduces three changes simultaneously: (1) switching from direct full-range regression to three-channel windowed regression targets, (2) the soft fusion operator (Eq. 3-4), and (3) a Transformer-based residual refinement module with additional parameters and self-attention. The improvement attributed to FRT could be driven by the added Transformer capacity rather than by the window decomposition itself. An additional ablation applying the Transformer residual refinement to direct full-range regression (same module, no windowing) would be needed.

Authors: The referee is correct that the current ablation does not isolate the contribution of windowed regression from that of the Transformer refinement module. We acknowledge this confound and will add the requested ablation: a Transformer-based residual refinement module applied to direct full-range regression (same Transformer architecture, no window decomposition, no soft fusion). This will allow us to disentangle the effect of the window prior from the effect of added Transformer capacity. We expect this experiment to show that the window decomposition contributes meaningfully beyond what the Transformer alone provides, but we agree that the claim must be supported empirically rather than assumed. We will revise Table 2 to include this additional row and update the discussion accordingly. revision: yes
Referee: Section 2.1, Eq. (3): The soft fusion operator derives fusion masks solely from the soft-tissue window prediction y(soft), assuming that the soft-tissue window's intensity bounds reliably characterize transitions between all three windows including lung and bone. If the predicted soft-tissue window has errors at its boundaries, the fusion masks for lung and bone regions will be incorrect, propagating artifacts into the final fused CT. No sensitivity analysis to soft-tissue prediction errors at window boundaries is provided.

Authors: This is a valid concern. The soft fusion operator indeed relies on the soft-tissue window prediction to derive the transition masks for all three windows, and errors at the soft-tissue window boundaries could propagate into the fused output. We will add a sensitivity analysis to the revised manuscript. Specifically, we plan to: (1) inject controlled perturbations into the soft-tissue window predictions at boundary regions and measure the resulting error in the fused CT, and (2) report the empirical distribution of soft-tissue window prediction errors at the transition boundaries on the test set, quantifying how often and how severely the boundary predictions deviate from ground truth. We note that the soft fusion function (Eq. 3-4) uses a transition band controlled by tau=0.2, which provides a smooth blending region rather than a hard boundary, offering some robustness to small prediction errors. However, the referee is correct that this robustness has not been quantified, and we will address that gap. revision: yes
Referee: Table 1(a): No statistical significance tests are reported for the metric improvements over baselines. The improvements over MedNeXt are meaningful in absolute terms, but the standard deviations are large. Without paired statistical tests, it is unclear whether these improvements are statistically significant.

Authors: We agree that paired statistical significance tests should be reported, especially given the large standard deviations in the per-case metrics. In the revised manuscript, we will conduct paired Wilcoxon signed-rank tests (or paired t-tests, depending on normality) comparing WING against each baseline on a per-case basis for all four metrics (MAE, MS-SSIM, PSNR, DICE) on both tasks. We will report p-values in Table 1(a) and discuss the statistical significance of the improvements. We note that the per-case standard deviations reported in the current manuscript reflect inter-patient variability (which is substantial due to the multi-anatomy, multi-center nature of SynthRAD2025), while the paired test will control for this by comparing methods on the same cases. The reduced standard deviation of WING (e.g., 20.77 vs. 23.53 for MedNeXt on MRI-to-CT MAE) is encouraging, but we agree that a formal paired test is necessary to substantiate the claimed improvements. revision: yes

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity found; one minor self-citation that is not load-bearing.

full rationale

The paper's derivation chain is self-contained. The CT window settings (WL/WW) are standard radiological values from [13] (radiopaedia.org), not fitted to the benchmark data. The soft fusion operator (Eq. 3-4) is a differentiable design with a single hyperparameter τ=0.2 set empirically, not a fitted constant that circularly defines the result. The evaluation uses the external SynthRAD2025 benchmark [23] with standard metrics (MAE, MS-SSIM, PSNR, DICE) computed following the challenge protocol. The architectural components (MedNeXt backbone [19], Inception decomposition [28], Transformer [25], PatchGAN [10]) are all externally cited and independently developed. Self-citation [12] (GANext by Mei, Xia, Fan) appears only in the related-work reference list and is not invoked as a load-bearing theorem, uniqueness result, or ansatz that would make the central claim circular. The skeptic's concern about the ablation confounding window decomposition with Transformer capacity (Table 2, +FRT row) is a legitimate experimental-design issue, but it concerns whether the evidence isolates the proposed mechanism, not whether the derivation reduces to its inputs by construction. No step in the paper's chain reduces to its own inputs by definition or by self-citation.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical entities, particles, forces, or dimensions are postulated. The method introduces architectural components (GIB, FRT, JDisc) but these are engineering constructs, not ontological commitments. All free parameters are standard hyperparameters or domain-established window settings.

free parameters (3)

τ (soft fusion hardness) = 0.2
Controls smoothness of fusion boundary transitions in Eq. 3-4. Set empirically without systematic sweep.
λ_win, λ_ct, λ_lpips, λ_adv = 1, 3, 0.4, 0.1
Loss weights in Eq. 7-8, empirically configured. No sensitivity analysis provided.
Window settings (WL/WW for lung, soft, bone) = (-600,800), (0,400), (1100,1800)
Fixed following radiological practice [13], not fitted to data. These are domain-standard values, reducing circularity burden.

assumptions (3)

domain assumption CT intensities are structure-deterministic and window-separable (Sec. 1, paragraph 5)
The entire framework depends on the assumption that fixed intensity ranges correspond deterministically to anatomical structures. This is well-established in radiology but is an inductive prior, not a proven theorem.
domain assumption Soft-tissue window bounds characterize transitions between all three windows (Sec. 2.1, Eq. 3)
The fusion masks for lung and bone are derived solely from the soft-tissue window prediction. This assumes the soft-tissue window's intensity range is a sufficient statistic for all window boundaries.
domain assumption Deformable registration via ConvexAdam provides accurate voxel-wise supervision (Sec. 3.1)
Training depends on paired MRI/CBCT-CT data with accurate voxel correspondence. Registration errors directly affect all learned metrics but are not quantified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of WING: A Window-Prior-Based Generative Network with Gated Inception for Cross-Modality CT Synthesis." pith.science (2026). https://pith.science/paper/ZDAQLNGN

@misc{pith2026260706234,
  author       = {Pith},
  title        = {Pith review of: WING: A Window-Prior-Based Generative Network with Gated Inception for Cross-Modality CT Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZDAQLNGN}},
  note         = {Machine review of arXiv:2607.06234}
}

read the original abstract

Generating CT volumes from MRI and CBCT can improve treatment planning in adaptive radiotherapy while avoiding additional radiation exposure. However, direct regression of CT intensities is challenged by the inherently high dynamic range and long-tailed distributions, thereby averaging out sparse yet clinically important structures. To alleviate this issue, we reformulate the regression target into multiple windowed representations, leveraging the inductive prior that CT intensities are structure-deterministic and window-separable. These windowed views exhibit smoother distributions and admit structured fusion back to the full-range CT. Building on this reformulation, we introduce WING, a WINdow-prior-based Generative network comprising: 1) a new Gated Inception Generator to produce multi-window predictions, enabling multi-shape kernel interactions to capture cross-modality correspondence; 2) a Fuse-and-Refine Transformer to aggregate the windowed outputs and learn residuals for detail refinement; and 3) a joint adversarial training objective to enhance window-conditioned realism. Extensive experiments demonstrate that our compact WING achieves state-of-the-art performance on the MRI-to-CT and CBCT-to-CT benchmarks, while supporting multi-anatomy synthesis with a single model.

Figures

Figures reproduced from arXiv: 2607.06234 by the authors.

**Figure 1.** Illustration of raw CT images and three non-overlapping windows of lung, soft tissue, and bone. The windowed views are obtained using the fixed window level (WL) and window width (WW), featuring more learnable distributions. for accurate proton and photon dose calculation [4,24]. To optimize treatment planning in MRI- or CBCT-only ART workflows, synthetic CT (sCT) has recently been introduced to provide CT-equivale… view at source ↗

**Figure 2.** Model overview of WING and block illustration of GIG and FRT. “DWConv” means depth-wise convolution [7], and “PWConv” means point-wise convolution [7]. “Soft fuse” denotes a differentiable operator defined in Eq. 4. The soft-tissue mask, defined as m(soft) = 1 − m(lung) − m(bone), forms a trapezoidal profile and takes the value of 1 for y(soft) ∈ [−τ, τ ]. The fused CT image is computed as the sum of weighted windo… view at source ↗

**Figure 3.** Visual comparison on MRI-to-CT and CBCT-to-CT tasks. A ×4 zoomed region is highlighted below. All gray values are windowed between -1000 HU and 1500 HU. 3.3 Performance Comparison We compare WING with representative GANs w/wo geometric guidance, as well as the top three models in the SynthRAD challenge. Following [23], we evaluate image fidelity using MAE, MS-SSIM, and PSNR, and assess geometric consistency via DICE… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages

[1]

Enhancing Cross-Modality Synthesis: Subvolume Merging for MRI-to-CT Conversion

Fan, F., Qiu, J., Huang, Y., Maier, A.: Enhancing cross-modality synthesis: Sub- volume merging for mri-to-ct conversion. arXiv preprint arXiv:2409.05982 (2024)

work page Pith review arXiv 2024
[2]

International Journal of Radiation Oncology* Biology* Physics109(4), 1054–1075 (2021)

Glide-Hurst, C.K., Lee, P., Yock, A.D., Olsen, J.R., Cao, M., Siddiqui, F., Parker, W., Doemer, A., Rong, Y., Kishan, A.U., et al.: Adaptive radiation therapy (art) strategies and technical considerations: a state of the art review from nrg oncology. International Journal of Radiation Oncology* Biology* Physics109(4), 1054–1075 (2021)

work page 2021
[3]

Medical Physics (2025)

Gong, C., Jian, J., Huang, Y., Luo, M., Ding, S., Yuan, X., Wang, X., Zhang, Y.: Boundary information-guided adversarial diffusion model for efficient unsupervised synthetic ct generation. Medical Physics (2025)

work page 2025
[4]

Grégoire, V., Mackie, T.R.: State of the art on dose prescription, reporting and recording in intensity-modulated radiation therapy (icru report no. 83). Can- cer/radiothérapie15(6-7), 555–559 (2011)

work page 2011
[5]

Bioengineering12(12), 1297 (2025)

Guo, Y., Luo, Y., Hooshangnejad, H., Zhang, R., Feng, X., Chen, Q., Ngwa, W., Ding, K.: Deep learning for ct synthesis in radiotherapy. Bioengineering12(12), 1297 (2025)

work page 2025
[6]

In: International MICCAI brainlesion workshop

Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.: Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In: International MICCAI brainlesion workshop. pp. 272–284. Springer (2021)

work page 2021
[7]

MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., An- dreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)

work page Pith review arXiv 2017
[8]

Medical image analysis97, 103276 (2024)

Huijben, E.M., Terpstra, M.L., Pai, S., Thummerer, A., Koopmans, P., Afonso, M., Van Eijnatten, M., Gurney-Champion, O., Chen, Z., Zhang, Y., et al.: Generating synthetic computed tomography for radiotherapy: Synthrad2023 challenge report. Medical image analysis97, 103276 (2024)

work page 2024

Show all 30 references

[9]

Nature methods18(2), 203–211 (2021)

Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods18(2), 203–211 (2021)

2021
[10]

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

Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with condi- tional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1125–1134 (2017)

2017
[11]

Nature Biomedical Engineering pp

Ma, L., Liang, H., He, Y., Wang, W., Yan, Z., Li, W., Wang, R., Li, Y., Lizhu, Y., Liu, Y., et al.: A vision–language pretrained transformer for versatile clinical respiratory disease applications. Nature Biomedical Engineering pp. 1–19 (2025)

2025
[12]

arXiv preprint arXiv:2512.19336 (2025) 10 S

Mei, S., Xia, Y., Fan, F.: Ganext: A fully convnext-enhanced generative adversarial network for mri- and cbct-to-ct synthesis. arXiv preprint arXiv:2512.19336 (2025) 10 S. Mei et al

2025
[13]

53347/rID-52108, https://radiopaedia.org/articles/52108, last revised: 2025-01-07

Murphy,A.,Feger,J.,Ismail,M.,etal.:Windowing(ct)(2017).https://doi.org/10. 53347/rID-52108, https://radiopaedia.org/articles/52108, last revised: 2025-01-07

2017
[14]

Advances in neural information processing sys- tems32(2019)

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. Advances in neural information processing sys- tems32(2019)

2019
[15]

Medical physics51(3), 1847–1859 (2024)

Peng, J., Qiu, R.L., Wynne, J.F., Chang, C.W., Pan, S., Wang, T., Roper, J., Liu, T., Patel, P.R., Yu, D.S., et al.: Cbct-based synthetic ct image generation using conditional denoising diffusion probabilistic model. Medical physics51(3), 1847–1859 (2024)

2024
[16]

In: International Conference on Medical Image Comput- ing and Computer-Assisted Intervention

Phan, V.M.H., Xie, Y., Zhang, B., Qi, Y., Liao, Z., Perperidis, A., Phung, S.L., Verjans, J.W., To, M.S.: Structural attention: Rethinking transformer for unpaired medical image synthesis. In: International Conference on Medical Image Comput- ing and Computer-Assisted Interven...

2024
[17]

arXiv preprint arXiv:2503.08515 (2025)

Poch, D.V., Estievenart, Y., Zhalieva, E., Patra, S., Yaqub, M., Taieb, S.B.: Segmentation-guided ct synthesis with pixel-wise conformal uncertainty bounds. arXiv preprint arXiv:2503.08515 (2025)

2025 arXiv
[18]

In: International Conference on Medical image computing and computer-assisted intervention

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234–241. Springer (2015)

2015
[19]

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

Roy, S., Koehler, G., Ulrich, C., Baumgartner, M., Petersen, J., Isensee, F., Jaeger, P.F., Maier-Hein, K.H.: Mednext: transformer-driven scaling of convnets for medi- cal image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Interve...

2023
[20]

IEEE Transactions on Medical Imaging (2024)

Siebert, H., Großbröhmer, C., Hansen, L., Heinrich, M.P.: Convexadam: Self- configuring dual-optimisation-based 3d multitask medical image registration. IEEE Transactions on Medical Imaging (2024)

2024
[21]

In: Seminars in radiation oncology

Sonke, J.J., Aznar, M., Rasch, C.: Adaptive radiotherapy for anatomical changes. In: Seminars in radiation oncology. vol. 29, pp. 245–257. Elsevier (2019)

2019
[22]

Medical physics48(11), 6537–6566 (2021)

Spadea, M.F., Maspero, M., Zaffino, P., Seco, J.: Deep learning based synthetic-ct generation in radiotherapy and pet: a review. Medical physics48(11), 6537–6566 (2021)

2021
[23]

Medical physics52(7), e17981 (2025)

Thummerer, A., van der Bijl, E., Galapon, A.J., Kamp, F., Savenije, M., Muijs, C., Aluwini, S., Steenbakkers, R.J., Beuel, S., Intven, M.P., et al.: Synthrad2025 grand challenge dataset: Generating synthetic cts for radiotherapy from head to abdomen. Medical physics52(7), e179...

2025
[24]

Medical physics48(12), 7673–7684 (2021)

Thummerer, A., Seller Oria, C., Zaffino, P., Meijers, A., Guterres Marmitt, G., Wijsman, R., Seco, J., Langendijk, J.A., Knopf, A.C., Spadea, M.F., et al.: Clinical suitability of deep learning based synthetic cts for adaptive proton therapy of lung cancer. Medical physics48(1...

2021
[25]

Advances in neural information pro- cessing systems30(2017)

Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017)

2017
[26]

Physics and Imaging in Radiation Oncology28, 100512 (2023)

Yoganathan, S., Aouadi, S., Ahmed, S., Paloor, S., Torfeh, T., Al-Hammadi, N., Hammoud,R.:Generatingsyntheticimagesfromconebeamcomputedtomography using self-attention residual unet for head and neck radiotherapy. Physics and Imaging in Radiation Oncology28, 100512 (2023)

2023
[27]

IEEE trans- actions on medical imaging38(7), 1750–1762 (2019) WING: A Window-Prior-Based Generative Network for CT Synthesis 11

Yu, B., Zhou, L., Wang, L., Shi, Y., Fripp, J., Bourgeat, P.: Ea-gans: edge-aware generative adversarial networks for cross-modality mr image synthesis. IEEE trans- actions on medical imaging38(7), 1750–1762 (2019) WING: A Window-Prior-Based Generative Network for CT Synthesis 11

2019
[28]

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

Yu, W., Zhou, P., Yan, S., Wang, X.: Inceptionnext: When inception meets con- vnext. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 5672–5683 (2024)

2024
[29]

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

Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586–595 (2018)

2018
[30]

In: Proceedings of the IEEE interna- tional conference on computer vision

Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE interna- tional conference on computer vision. pp. 2223–2232 (2017)

2017

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.

[1] [1]

Enhancing Cross-Modality Synthesis: Subvolume Merging for MRI-to-CT Conversion

Fan, F., Qiu, J., Huang, Y., Maier, A.: Enhancing cross-modality synthesis: Sub- volume merging for mri-to-ct conversion. arXiv preprint arXiv:2409.05982 (2024)

work page Pith review arXiv 2024

[2] [2]

International Journal of Radiation Oncology* Biology* Physics109(4), 1054–1075 (2021)

Glide-Hurst, C.K., Lee, P., Yock, A.D., Olsen, J.R., Cao, M., Siddiqui, F., Parker, W., Doemer, A., Rong, Y., Kishan, A.U., et al.: Adaptive radiation therapy (art) strategies and technical considerations: a state of the art review from nrg oncology. International Journal of Radiation Oncology* Biology* Physics109(4), 1054–1075 (2021)

work page 2021

[3] [3]

Medical Physics (2025)

Gong, C., Jian, J., Huang, Y., Luo, M., Ding, S., Yuan, X., Wang, X., Zhang, Y.: Boundary information-guided adversarial diffusion model for efficient unsupervised synthetic ct generation. Medical Physics (2025)

work page 2025

[4] [4]

Grégoire, V., Mackie, T.R.: State of the art on dose prescription, reporting and recording in intensity-modulated radiation therapy (icru report no. 83). Can- cer/radiothérapie15(6-7), 555–559 (2011)

work page 2011

[5] [5]

Bioengineering12(12), 1297 (2025)

Guo, Y., Luo, Y., Hooshangnejad, H., Zhang, R., Feng, X., Chen, Q., Ngwa, W., Ding, K.: Deep learning for ct synthesis in radiotherapy. Bioengineering12(12), 1297 (2025)

work page 2025

[6] [6]

In: International MICCAI brainlesion workshop

Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.: Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In: International MICCAI brainlesion workshop. pp. 272–284. Springer (2021)

work page 2021

[7] [7]

MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., An- dreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)

work page Pith review arXiv 2017

[8] [8]

Medical image analysis97, 103276 (2024)

Huijben, E.M., Terpstra, M.L., Pai, S., Thummerer, A., Koopmans, P., Afonso, M., Van Eijnatten, M., Gurney-Champion, O., Chen, Z., Zhang, Y., et al.: Generating synthetic computed tomography for radiotherapy: Synthrad2023 challenge report. Medical image analysis97, 103276 (2024)

work page 2024