Volumetric Directional Diffusion: Anchoring Uncertainty Quantification in Anatomical Consensus for Ambiguous Medical Image Segmentation

Chao Wu; David Doermann; Kangxian Xie; Mahesh Bhosale; Mingchen Gao; Pouya Karimian

REVIEW 3 major objections 6 minor 27 references

This paper claims that anchoring a 3D diffusion model to a coarse anatomical prior turns generation into boundary-residual exploration, yielding diverse expert-plausible segmentations without breaking volumetric structure.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 18:57 UTC pith:WYC6EMRN

load-bearing objection A practical anchored-diffusion refiner for ambiguous 3D segmentation with a clear plug-in value, but the reverse step is asserted rather than derived and needs fixing before the UQ claims hold. the 3 major comments →

arxiv 2603.04024 v2 pith:WYC6EMRN submitted 2026-03-04 cs.CV cs.AI

Volumetric Directional Diffusion: Anchoring Uncertainty Quantification in Anatomical Consensus for Ambiguous Medical Image Segmentation

Chao Wu , Mahesh Bhosale , Kangxian Xie , Pouya Karimian , David Doermann , Mingchen Gao This is my paper

classification cs.CV cs.AI

keywords medical image segmentationdiffusion modelsuncertainty quantificationaleatoric uncertaintyanatomical anchoringboundary residual explorationvolumetric consistencymulti-rater segmentation

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tackles the fidelity-diversity trade-off in ambiguous 3D medical segmentation. It proposes Volumetric Directional Diffusion (VDD), which anchors the diffusion trajectory to a coarse consensus mask so that generation is limited to plausible boundary variations around that mask instead of recreating whole organs from noise. The authors claim this preserves slice-to-slice volumetric consistency while achieving state-of-the-art uncertainty maps on three multi-rater datasets, with accuracy close to deterministic baselines. The core message is that stochastic generative models become clinically useful for inter-observer variability when their randomness is constrained to anatomically meaningful degrees of freedom.

Core claim

On the paper's own terms, the central discovery is that a coarse deterministic segmentation prior can be woven into a diffusion process as a continuous spatial bias, so the forward process drifts toward the prior rather than isotropic noise and the reverse process only has to refine the residual boundary field. The expected terminal state of the forward chain is exactly the prior (Eq. 6), which means the model's generative search space is a structured neighborhood around the consensus mask. Across three multi-rater datasets, VDD reports better Generalized Energy Distance, Collective Insight, and spatial cross-correlation than VAE-based and 2D diffusion baselines, with Dice scores near determ

What carries the argument

The central mechanism is the directional forward transition y_t = sqrt(alpha_t) y_{t-1} + (1 - sqrt(alpha_t)) y_hat + noise, with y_hat a coarse anatomical prior, and its reverse counterpart that plugs the predicted clean boundary y0_theta into the anchored marginal. This 'anatomical anchoring' forces the Markov chain to stay in a bracket around y_hat at every step, so the generative task becomes residual exploration of boundary ambiguity rather than full-mask synthesis. The expectation identity lim E[y_t] = y_hat as alpha_bar_T goes to 0 is the formal anchor.

Load-bearing premise

The load-bearing premise is that the reverse sampling step (Eq. 8) is the correct inverse of the anchored forward process, but the paper never derives this posterior; if that step mismatches the forward marginal, VDD samples are not true posterior draws and the uncertainty maps are not well-defined.

What would settle it

Derive q(y_{t-1} | y_t, y0, y_hat) from the forward transition of Eq. (3) and compare with Eq. (8); if the two differ, run VDD on a synthetic dataset with a known ground-truth expert-label distribution and measure whether the sample distribution recovers the true inter-rater spread (e.g., via maximum mean discrepancy).

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

If VDD's results hold, uncertainty maps for ambiguous lesions can be derived from diffusion samples that stay within anatomically coherent topology, making them usable for radiation-planning or margin-assessment decisions.
VDD can act as a plug-and-play refiner on top of an existing deterministic segmentation network, improving uncertainty calibration without retraining the base model.
The anchored process requires fewer denoising steps (T=50) than typical 2D diffusion baselines, bringing generative uncertainty estimation closer to clinical latency constraints.
Reported improvements in CI and GED indicate that VDD samples align better with true inter-rater disagreement regions than VAE- or 2D-diffusion-based stochastic segmentation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the reverse step of Eq. (8) is a faithful inverse of Eq. (3), VDD can be viewed as a form of controlled generation where the anchor acts as a learned condition; the same anchoring trick could apply to other structured-output tasks such as 3D object reconstruction or label refinement.
The paper leaves the reverse posterior un-derived; a formal derivation and a check on whether the sampler matches the posterior would be the next natural step to make the uncertainty maps mathematically well-defined.
Because the anchor is a deterministic network, VDD might inherit the base model's biases; testing VDD with different anchor models could reveal how much of the uncertainty is anatomical versus inherited.
A testable extension is to make the anchor itself stochastic (e.g., sampling from a distribution of plausible cores) and compare sample diversity against VDD's fixed-anchor variant.

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

A practical anchored-diffusion refiner for ambiguous 3D segmentation with a clear plug-in value, but the reverse step is asserted rather than derived and needs fixing before the UQ claims hold.

read the letter

What to know: VDD is a plausible plug-in for turning a deterministic 3D segmentation network into an uncertainty-aware sampler. The reported gains on GED/CI over 2D diffusion and VAE baselines across three datasets are consistent with the core story: anchor the generative trajectory to a coarse consensus prior and let diffusion explore only boundary residuals. That is a sensible way to sidestep the topology-collapse problem of full-mask 3D diffusion.

The paper is honest about the prior being imperfect, and the experiments cover LIDC-IDRI, KiTS21, and ISBI 2015 with a held-out split. The ablation separating anchored from Gaussian latent sampling is useful. The mechanism is a direct transplant of Hou et al.'s directional label diffusion, but the paper cites that work and the empirical demonstration in the 3D medical segmentation setting is new.

Now the soft spots, in proportion. The stress-test note is right: Eq. (8) does not follow from Eq. (3). The forward marginal (Eq. 4) is a standard DDPM on the residual z = y - y_hat. The correct posterior q(y_{t-1}|y_t,y0,y_hat) would contain a term depending on z_t (equivalently on eps_theta) beyond what Eq. (8) writes. Without defining sigma_t and providing the derivation, the sampler is an unvalidated kernel, and the uncertainty maps are not anchored to a learned posterior. The missing training objective in Section 2.3 compounds the problem — we never see the loss that would define the learned noise prediction. Also: a duplicated Section 2.2, a citation placeholder "[cite: 30]," no error bars or significance tests, no code or data, and the latency comparison in Table 3 uses per-slice time for 2D models and per-volume time for 3D models, which is not apples-to-apples. The "Proof of Expected Convergence" (Eqs. 5-6) shows only convergence of the mean; it is a sanity check, not a distributional or topological guarantee.

That said, the central idea is not invalid. The empirical gains are plausible and the experiments are reasonably designed. The missing derivation is a fixable gap, not a fatal flaw — assuming the intended reverse step can be derived, or a standard DDPM posterior on the residual is used instead.

Who this is for: groups working on uncertainty quantification in medical segmentation who want a lightweight alternative to full 3D diffusion. A serious referee should see it, but the paper needs major revision before acceptance.

Referee Report

3 major / 6 minor

Summary. The paper proposes Volumetric Directional Diffusion (VDD), a 3D diffusion model for ambiguous medical image segmentation that anchors the generative process to a coarse anatomical prior y_hat. The forward process in Eq. (3) drifts noisy masks toward y_hat instead of isotropic zero, and the reverse process in Eq. (8) is designed to generate plausible boundary variations around y_hat while preserving 3D structure. The paper claims that this 'residual exploration' framework improves uncertainty metrics (GED, CI, SNCC) and remains competitive in Dice/HD95 on three multi-rater datasets (LIDC-IDRI, KiTS21, ISBI 2015). The core technical claim is that VDD's anchored diffusion yields a safe, topologically consistent sampling space and that the resulting uncertainty maps are calibrated to expert disagreement.

Significance. If the derivation were sound, VDD would be a valuable contribution: it reframes diffusion-based medical segmentation as residual exploration around a deterministic consensus prior, directly addressing the 3D consistency gap of slice-wise 2D diffusion models. The paper evaluates on three public datasets, includes an ablation of the anchor and the number of samples, and reports inference cost, which are useful practical elements. However, the central sampling equation is asserted rather than derived, the training objective is never written, and the convergence argument only concerns the expectation, not the distribution. As a result, the experimental numbers in Table 1 cannot currently be interpreted as evidence for the uncertainty-calibration claims. The idea is promising but the mathematical foundation must be repaired before the claims can be assessed.

major comments (3)

[§2.3, Eq. (8)] The reverse sampling step is not derived from the forward process in Eq. (3). Let z_t = y_t - y_hat; then Eq. (3) is the standard DDPM forward noising z_t = sqrt(alpha_t) z_{t-1} + sqrt(1-alpha_t) epsilon_t, whose posterior is z_{t-1} = a_t z_t + b_t z_0 + sigma_t eps with a_t = sqrt(alpha_t)(1 - alpha_bar_{t-1})/(1 - alpha_bar_t) and b_t = sqrt(alpha_bar_{t-1})(1 - alpha_t)/(1 - alpha_bar_t). Substituting Eq. (7) into Eq. (8) gives z_{t-1} = sqrt(alpha_bar_{t-1}/alpha_bar_t) z_t - sqrt(alpha_bar_{t-1}(1-alpha_bar_t)/alpha_bar_t) eps_theta + sigma_t eps, whose z_t coefficient is 1/sqrt(alpha_t), not a_t. These two forms coincide only for a special choice of schedule and noise scale that is never specified. Therefore samples from Eq. (8) are not shown to be samples from the time-reversal of Eq. (3), and the uncertainty maps in Table 1 are not anchored to any well-defined target distributi
[§2.3, after Eq. (7)] The training objective that defines the learned posterior is never written. The text mentions a 'boundary-aware objective' but gives no equation for the loss on eps_theta (or on y0_theta). Without this, the network's predicted clean mask has no defined optimality criterion, and the reverse process cannot be connected to a learned approximation of the forward posterior. This missing equation is load-bearing because the uncertainty claims depend on the distribution induced by the trained denoiser.
[§2.2, Eqs. (5)-(6)] The proof of 'Expected Convergence' is not sufficient for the claims drawn from it. First, Eq. (5) writes sqrt(alpha_t) instead of sqrt(alpha_bar_t); the expectation of Eq. (4) is sqrt(alpha_bar_t) y0 + (1 - sqrt(alpha_bar_t)) y_hat. Second, even after correcting this typo, the argument only shows that E[y_T] = y_hat when alpha_bar_T = 0. From Eq. (4), the marginal variance at T is (1 - alpha_bar_T) I = I, so y_T is an isotropic Gaussian centered at y_hat, not a sample in a 'bounded neighborhood' or 'safe search space'. Expectation does not imply that individual trajectories stay near y_hat, nor does it provide any topological or slice-consistency guarantee. The claimed mitigation of fractures and hallucinations is therefore not established by the provided derivation.

minor comments (6)

[§2.1 and §2.2] The two subsections are essentially identical; the manuscript appears to contain a duplicated block. Please remove one and renumber the equations consistently.
[Table 1] The text states that VDD 'significantly improves' GED and CI, but no error bars, confidence intervals, or statistical tests are reported. Without them, the claimed significance is not supported.
[Fig. 1 and Fig. 3] Several symbols in Fig. 1 are unexplained ('S', the concatenation symbol, the sigmoid/segmentation branch), and the label 'CSSigmoidConcatenation' is garbled. In Fig. 3, the abbreviation 'Prob U.Gaussian' is unclear.
[§3.4, Table 3] The inference-time comparison mixes per-slice times for 2D models and per-volume times for 3D models. To be fair, report the total wall-clock time per test volume for all methods, including the cost of assembling 2D slices.
[References and formatting] There is a placeholder citation '[cite: 30]' in the introduction. Several references are incomplete or inconsistently formatted (e.g., missing page numbers, preprint identifiers). Please check the bibliography.
[§2.2, evaluation metrics] GED and CI are referenced only by citations; their definitions should be stated explicitly so that the reader can interpret the quantitative claims in Table 1.

Circularity Check

0 steps flagged

No construction-level circularity: the anchor is an independent network output, not a fitted parameter; the reverse-sampler gap in §2.3 is an omitted derivation (correctness risk) rather than a reduction to inputs.

full rationale

The derivation chain does not reduce any claimed prediction to its inputs. The coarse prior y_hat is described as 'a straightforward, coarse anatomical prior generated by a standard baseline network' (§2.2); it is not a parameter fitted to GED/CI or to the sampled uncertainty. Eq. (6) gives lim E[y_t]=y_hat only because the schedule is constrained to alpha_bar_T=0; this is a designed property of the forward chain, not a fitted prediction. The reverse step Eq. (8) is asserted without deriving q(y_{t-1}|y_t,y0,y_hat) from Eq. (3); substituting Eq. (7) into Eq. (8) yields a residual update that differs from the standard DDPM/DDIM posterior (it omits the z_t-dependent eps_theta term unless sigma_t is specially chosen). This is an omitted proof and a potential correctness flaw, but it is not circularity: Eq. (8) is not equal to Eq. (3) or Eq. (4) by construction, and no fitted value is renamed as a prediction. Evaluation is against external multi-rater benchmarks (LIDC-IDRI, KiTS21, ISBI 2015) with standard metrics. The only self-citation is [22] (Wu and Gao) in the motivational sentence 'recovering complex volumetric topology from pure noise is an extremely unconstrained inverse problem [8,22]'; it is not load-bearing, and the central method cites external work [11] for directional diffusion. Thus the paper is self-contained with respect to circularity; the reverse-sampler issue should be treated as a correctness/validation concern, not a circularity score increase.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new physical or ontological entities are introduced; the method is a generative process on existing voxel spaces. The ledger is dominated by the unproved reverse-step validity and the dependence on the anchor prior, plus unspecified schedule and noise hyperparameters.

free parameters (4)

Anatomical anchor prior y_hat = output of a deterministic 3D U-Net/nnU-Net trained on consensus labels
The entire generative trajectory is pulled toward this learned map; its quality defines the search space for boundary uncertainty.
Diffusion noise schedule alpha_t = not specified in the paper
Controls how quickly the trajectory approaches the anchor; no formula or hyperparameter values are given.
Number of diffusion steps T = 50
Chosen for inference speed; no ablation over T is reported.
Reverse-step noise sigma_t (Eq. 8) = unspecified
The noise scale in the reverse update is not defined; it may be beta_bar_{t-1} but this is never stated.

axioms (4)

ad hoc to paper The reverse update in Eq. (8) correctly inverts the forward process in Eq. (3).
No posterior derivation is given; this is the load-bearing unproved step of the generative sampler.
domain assumption The consensus prior y_hat is a good enough macroscopic anatomy that all clinically plausible expert contours lie near it.
If the anchor is far from real expert contours, residual exploration cannot recover the missing uncertainty.
domain assumption Treating a binary segmentation as a continuous value y0 in [-1,1]^DxHxW is a valid relaxation for the diffusion objective.
The paper does not discuss how the continuous-valued diffusion output is binarized for metrics such as Dice/HD95.
ad hoc to paper The limiting expectation E[y_T] = y_hat implies a 'safe search space' with preserved topology.
Equality of expectations does not guarantee distributional concentration or topological stability; this is used to justify the method's core claim.

pith-pipeline@v1.3.0-alltime-deepseek · 7379 in / 16893 out tokens · 148119 ms · 2026-08-02T18:57:02.562665+00:00 · methodology

0 comments

read the original abstract

Ambiguous 3D medical image segmentation often involves boundaries where different expert delineations are non-identical yet clinically plausible. Modeling such inter-observer variability requires a careful balance between diversity and anatomical fidelity: deterministic models preserve coherent volumetric structures but collapse expert disagreement into a single mask, while stochastic generative models can produce diverse samples but may introduce disconnected components or slice-to-slice inconsistency when generating full 3D masks from unstructured noise. We propose Volumetric Directional Diffusion (VDD), a prior-anchored diffusion framework that shifts stochastic generation from full-mask synthesis to residual boundary exploration. VDD uses a coarse consensus prediction as an anatomical anchor and learns a directional diffusion process to generate plausible boundary variations around ambiguous regions while preserving stable volumetric topology. Experiments on three multi-rater datasets, including LIDC-IDRI, KiTS21, and ISBI 2015, show that VDD improves uncertainty distribution alignment while maintaining competitive segmentation accuracy and 3D structural consistency. These results suggest that prior-anchored residual diffusion can model clinically plausible expert disagreement without sacrificing anatomical fidelity.

Figures

Figures reproduced from arXiv: 2603.04024 by Chao Wu, David Doermann, Kangxian Xie, Mahesh Bhosale, Mingchen Gao, Pouya Karimian.

**Figure 1.** Figure 1: Overview of the proposed Volumetric Directional Diffusion (VDD) framework. The pipeline consists of three main stages: (1) Anatomical Anchoring: A deterministic 3D U-Net is trained on the consensus of divergent expert annotations to extract a coarse anatomical prior yˆ. (2) Directional Forward Process: Unlike the original DDPM that diffuses towards isotropic Gaussian noise, VDD anchors the forward trajec… view at source ↗

**Figure 2.** Figure 2: 3D visual comparison of uncertainty quantification on highly ambiguous LIDCIDRI nodules. From left to right, the four blocks represent: (1) Ground Truth, (2) CCDM, (3) DiffOSeg, and (4) VDD (Ours). 3.2 Results Balancing Anatomical Fidelity and Generative Diversity As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Visual comparison of uncertainty maps among VDD, Prob Unet and VDD(Gaussian) (CCDM, DiffOSeg) lack volumetric constraints, reconstructing morphology independently and disrupting topology. We analyze this across four morphotypes: (a) Dumbbell: 2D baselines merge distinct foci into a monolithic “peanut” shape, whereas VDD preserves the delicate bridging topology. (b) Pleural-tail & (d) Vessel-attached: 2D m… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

27 extracted references · 1 canonical work pages

[1]

arXiv preprint arXiv:2112.00390 (2021)

Amit, T., Eliashiv, T., Dekel, T., et al.: Segdiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390 (2021)

Pith/arXiv arXiv 2021
[2]

Medical physics38(2), 915–931 (2011)

Armato III, S.G., McLennan, G., Bidaut, L., McNitt-Gray, M.F., Meyer, C.R., Reeves, A.P., Zhao, B., Aberle, D.R., Henschke, C.I., Hoffman, E.A., et al.: The lungimagedatabaseconsortium(lidc)andimagedatabaseresourceinitiative(idri): a completed reference database of lung nodules on ct scans. Medical physics38(2), 915–931 (2011)

2011
[3]

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

Baru, L.B., Dadi, K., Chakraborti, T., Bapi, R.S.: Ambiguous medical image seg- mentation using diffusion schrödinger bridge. In: International Conference on Med- ical Image Computing and Computer-Assisted Intervention. pp. 25–35. Springer (2025)

2025
[4]

NeuroImage148, 77–102 (2017)

Carass, A., Roy, S., Jog, A., Cuzzocreo, J.L., Magrath, E., Gherman, A., Button, J., Nguyen, J., Prados, F., Sudre, C.H., et al.: Longitudinal multiple sclerosis lesion segmentation: resource and challenge. NeuroImage148, 77–102 (2017)

2017
[5]

In: 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI)

Chen, R., Zhang, X., Lin, T., Yu, S.: Emssd: Two-stage model enhancing medical image segmentation based on stable diffusion. In: 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI). pp. 1–5. IEEE (2025)

2025
[6]

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

Choo, K., Jun, Y., Yun, M., Hwang, S.J.: Slice-consistent 3d volumetric brain ct-to-mri translation with 2d brownian bridge diffusion model. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 657–667. Springer (2024)

2024
[7]

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

Fan, L., Zhang, M., Wang, Y., Lu, W., Yin, H.: Noisy label refinement based on discrete diffusion process in 3d ossicle segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 423–432. Springer (2025)

2025
[8]

Friedrich, P., Wolleb, J., Bieder, F., Durrer, A., Cattin, P.C.: Wdm: 3d wavelet diffusionmodelsforhigh-resolutionmedicalimagesynthesis.In:MICCAIworkshop on deep generative models. pp. 11–21. Springer (2024)

2024
[9]

arXiv preprint arXiv:2307.01984 (2023)

Heller, N., Isensee, F., Trofimova, D., Tejpaul, R., Zhao, Z., Chen, H., Wang, L., Golts, A., Khapun, D., Shats, D., et al.: The kits21 challenge: Automatic segmen- tation of kidneys, renal tumors, and renal cysts in corticomedullary-phase ct. arXiv preprint arXiv:2307.01984 (2023)

Pith/arXiv arXiv 2023
[10]

In: Advances inNeuralInformationProcessingSystems(NeurIPS).vol.33,pp.6840–6851(2020)

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Advances inNeuralInformationProcessingSystems(NeurIPS).vol.33,pp.6840–6851(2020)

2020
[11]

Hou, S., Jiang, G., Zhang, J., Yang, S., Guo, H., Guo, Y., Wang, W.: Directional la- beldiffusionmodelforlearningfromnoisylabels.In:ProceedingsoftheIEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25738–25748 (2025)

2025
[12]

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
[13]

arXiv preprint arXiv:2511.14063 (2025)

Jin, D., Xu, R., Zeng, J., Lan, R., Bai, Y., Sun, L., Chu, X.: Semantic con- text matters: Improving conditioning for autoregressive models. arXiv preprint arXiv:2511.14063 (2025)

arXiv 2025
[14]

Advances in neural information processing sys- tems31(2018) 10 C

Kohl, S., Romera-Paredes, B., Meyer, C., De Fauw, J., Ledsam, J.R., Maier-Hein, K., Eslami, S., Jimenez Rezende, D., Ronneberger, O.: A probabilistic u-net for segmentation of ambiguous images. Advances in neural information processing sys- tems31(2018) 10 C. Wu et al

2018
[15]

Proceedings of the AAAI Conference on Artificial In- telligence39(6), 5946–5954 (Apr 2025)

Lyu, Y., Cha, S.J., Jiang, C., Chowdury, A.Z., Hou, X., Harake, E.S., Kondepudi, A., Freudiger, C., Lee, H., Hollon, T.C.: Step-calibrated diffusion for biomedical optical image restoration. Proceedings of the AAAI Conference on Artificial In- telligence39(6), 5946–5954 (Apr 2025). https://doi.org/10.1609/aaai.v39i6.32635, https://ojs.aaai.org/index.php/A...

work page doi:10.1609/aaai.v39i6.32635 2025
[16]

Nature methods21(2), 195–212 (2024)

Maier-Hein, L., Reinke, A., Godau, P., Tizabi, M.D., Buettner, F., Christodoulou, E., Glocker, B., Isensee, F., Kleesiek, J., Kozubek, M., et al.: Metrics reloaded: recommendations for image analysis validation. Nature methods21(2), 195–212 (2024)

2024
[17]

In: Inter- national Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging

Nguyen, V.P., Trinh, N.H., Nguyen, D.M.L., Nguyen, P.L., Tran, Q.L.: Aleatoric uncertainty medical image segmentation estimation via flow matching. In: Inter- national Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging. pp. 134–144. Springer (2025)

2025
[18]

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

Rahman,A.,Valanarasu,J.M.J.,Hacihaliloglu,I.,Patel,V.M.:Ambiguousmedical image segmentation using diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11536–11546 (2023)

2023
[19]

Sha, T., Chen, Z., Cheng, Z., Zhai, H., Ding, X., Wang, K.: Fast- cad: A fairness-aware framework for non-contact stroke diagnosis (2025), https://arxiv.org/abs/2511.08887

Pith/arXiv arXiv 2025
[20]

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

Tivnan, M., Yoon, S., Chen, Z., Li, X., Wu, D., Li, Q.: Hallucination index: An image quality metric for generative reconstruction models. In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention. pp. 449–458. Springer (2024)

2024
[21]

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

Wolleb, J., Bieder, F., Sandkühler, R., Cattin, P.C.: Diffusion models for medical anomaly detection. In: International Conference on Medical image computing and computer-assisted intervention. pp. 35–45. Springer (2022)

2022
[22]

In: Medical Imaging with Deep Learning

Wu, C., Gao, M.: Simplex-aligned diffusion with cross-granularity interaction for robust medical image classification. In: Medical Imaging with Deep Learning
[23]

arXiv preprint arXiv:2507.20973 (2025)

Wu, C., Wang, Z., Xie, K., Devulapally, N.K., Lokhande, V.S., Gao, M.: Model- agnostic gender bias control for text-to-image generation via sparse autoencoder. arXiv preprint arXiv:2507.20973 (2025)

arXiv 2025
[24]

In: Medical imaging with deep learning

Wu, J., Fu, R., Fang, H., Zhang, Y., Yang, Y., Xiong, H., Liu, H., Xu, Y.: Med- segdiff: Medical image segmentation with diffusion probabilistic model. In: Medical imaging with deep learning. pp. 1623–1639. PMLR (2024)

2024
[25]

In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision

Zbinden, L., Doorenbos, L., Pissas, T., Huber, A.T., Sznitman, R., Márquez-Neila, P.: Stochastic segmentation with conditional categorical diffusion models. In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1119–1129 (2023)

2023
[26]

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

Zhang, H., Luo, X., Chen, Y., Li, K.: Diffoseg: Omni medical image segmenta- tion via multi-expert collaboration diffusion model. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 128–138. Springer (2025)

2025
[27]

Bioinformatics40(4), btae118 (2024)

Zhang,R.,Wu,C.,Yang,Q.,Liu,C.,Wang,Y.,Li,K.,Huang,L.,Zhou,F.:Molfes- cue: enhancing molecular property prediction in data-limited and imbalanced con- texts using few-shot and contrastive learning. Bioinformatics40(4), btae118 (2024)

2024

[1] [1]

arXiv preprint arXiv:2112.00390 (2021)

Amit, T., Eliashiv, T., Dekel, T., et al.: Segdiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390 (2021)

Pith/arXiv arXiv 2021

[2] [2]

Medical physics38(2), 915–931 (2011)

Armato III, S.G., McLennan, G., Bidaut, L., McNitt-Gray, M.F., Meyer, C.R., Reeves, A.P., Zhao, B., Aberle, D.R., Henschke, C.I., Hoffman, E.A., et al.: The lungimagedatabaseconsortium(lidc)andimagedatabaseresourceinitiative(idri): a completed reference database of lung nodules on ct scans. Medical physics38(2), 915–931 (2011)

2011

[3] [3]

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

Baru, L.B., Dadi, K., Chakraborti, T., Bapi, R.S.: Ambiguous medical image seg- mentation using diffusion schrödinger bridge. In: International Conference on Med- ical Image Computing and Computer-Assisted Intervention. pp. 25–35. Springer (2025)

2025

[4] [4]

NeuroImage148, 77–102 (2017)

Carass, A., Roy, S., Jog, A., Cuzzocreo, J.L., Magrath, E., Gherman, A., Button, J., Nguyen, J., Prados, F., Sudre, C.H., et al.: Longitudinal multiple sclerosis lesion segmentation: resource and challenge. NeuroImage148, 77–102 (2017)

2017

[5] [5]

In: 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI)

Chen, R., Zhang, X., Lin, T., Yu, S.: Emssd: Two-stage model enhancing medical image segmentation based on stable diffusion. In: 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI). pp. 1–5. IEEE (2025)

2025

[6] [6]

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

Choo, K., Jun, Y., Yun, M., Hwang, S.J.: Slice-consistent 3d volumetric brain ct-to-mri translation with 2d brownian bridge diffusion model. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 657–667. Springer (2024)

2024

[7] [7]

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

Fan, L., Zhang, M., Wang, Y., Lu, W., Yin, H.: Noisy label refinement based on discrete diffusion process in 3d ossicle segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 423–432. Springer (2025)

2025

[8] [8]

Friedrich, P., Wolleb, J., Bieder, F., Durrer, A., Cattin, P.C.: Wdm: 3d wavelet diffusionmodelsforhigh-resolutionmedicalimagesynthesis.In:MICCAIworkshop on deep generative models. pp. 11–21. Springer (2024)

2024

[9] [9]

arXiv preprint arXiv:2307.01984 (2023)

Heller, N., Isensee, F., Trofimova, D., Tejpaul, R., Zhao, Z., Chen, H., Wang, L., Golts, A., Khapun, D., Shats, D., et al.: The kits21 challenge: Automatic segmen- tation of kidneys, renal tumors, and renal cysts in corticomedullary-phase ct. arXiv preprint arXiv:2307.01984 (2023)

Pith/arXiv arXiv 2023

[10] [10]

In: Advances inNeuralInformationProcessingSystems(NeurIPS).vol.33,pp.6840–6851(2020)

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Advances inNeuralInformationProcessingSystems(NeurIPS).vol.33,pp.6840–6851(2020)

2020

[11] [11]

Hou, S., Jiang, G., Zhang, J., Yang, S., Guo, H., Guo, Y., Wang, W.: Directional la- beldiffusionmodelforlearningfromnoisylabels.In:ProceedingsoftheIEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25738–25748 (2025)

2025

[12] [12]

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

[13] [13]

arXiv preprint arXiv:2511.14063 (2025)

Jin, D., Xu, R., Zeng, J., Lan, R., Bai, Y., Sun, L., Chu, X.: Semantic con- text matters: Improving conditioning for autoregressive models. arXiv preprint arXiv:2511.14063 (2025)

arXiv 2025

[14] [14]

Advances in neural information processing sys- tems31(2018) 10 C

Kohl, S., Romera-Paredes, B., Meyer, C., De Fauw, J., Ledsam, J.R., Maier-Hein, K., Eslami, S., Jimenez Rezende, D., Ronneberger, O.: A probabilistic u-net for segmentation of ambiguous images. Advances in neural information processing sys- tems31(2018) 10 C. Wu et al

2018

[15] [15]

Proceedings of the AAAI Conference on Artificial In- telligence39(6), 5946–5954 (Apr 2025)

Lyu, Y., Cha, S.J., Jiang, C., Chowdury, A.Z., Hou, X., Harake, E.S., Kondepudi, A., Freudiger, C., Lee, H., Hollon, T.C.: Step-calibrated diffusion for biomedical optical image restoration. Proceedings of the AAAI Conference on Artificial In- telligence39(6), 5946–5954 (Apr 2025). https://doi.org/10.1609/aaai.v39i6.32635, https://ojs.aaai.org/index.php/A...

work page doi:10.1609/aaai.v39i6.32635 2025

[16] [16]

Nature methods21(2), 195–212 (2024)

Maier-Hein, L., Reinke, A., Godau, P., Tizabi, M.D., Buettner, F., Christodoulou, E., Glocker, B., Isensee, F., Kleesiek, J., Kozubek, M., et al.: Metrics reloaded: recommendations for image analysis validation. Nature methods21(2), 195–212 (2024)

2024

[17] [17]

In: Inter- national Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging

Nguyen, V.P., Trinh, N.H., Nguyen, D.M.L., Nguyen, P.L., Tran, Q.L.: Aleatoric uncertainty medical image segmentation estimation via flow matching. In: Inter- national Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging. pp. 134–144. Springer (2025)

2025

[18] [18]

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

Rahman,A.,Valanarasu,J.M.J.,Hacihaliloglu,I.,Patel,V.M.:Ambiguousmedical image segmentation using diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11536–11546 (2023)

2023

[19] [19]

Sha, T., Chen, Z., Cheng, Z., Zhai, H., Ding, X., Wang, K.: Fast- cad: A fairness-aware framework for non-contact stroke diagnosis (2025), https://arxiv.org/abs/2511.08887

Pith/arXiv arXiv 2025

[20] [20]

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

Tivnan, M., Yoon, S., Chen, Z., Li, X., Wu, D., Li, Q.: Hallucination index: An image quality metric for generative reconstruction models. In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention. pp. 449–458. Springer (2024)

2024

[21] [21]

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

Wolleb, J., Bieder, F., Sandkühler, R., Cattin, P.C.: Diffusion models for medical anomaly detection. In: International Conference on Medical image computing and computer-assisted intervention. pp. 35–45. Springer (2022)

2022

[22] [22]

In: Medical Imaging with Deep Learning

Wu, C., Gao, M.: Simplex-aligned diffusion with cross-granularity interaction for robust medical image classification. In: Medical Imaging with Deep Learning

[23] [23]

arXiv preprint arXiv:2507.20973 (2025)

Wu, C., Wang, Z., Xie, K., Devulapally, N.K., Lokhande, V.S., Gao, M.: Model- agnostic gender bias control for text-to-image generation via sparse autoencoder. arXiv preprint arXiv:2507.20973 (2025)

arXiv 2025

[24] [24]

In: Medical imaging with deep learning

Wu, J., Fu, R., Fang, H., Zhang, Y., Yang, Y., Xiong, H., Liu, H., Xu, Y.: Med- segdiff: Medical image segmentation with diffusion probabilistic model. In: Medical imaging with deep learning. pp. 1623–1639. PMLR (2024)

2024

[25] [25]

In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision

Zbinden, L., Doorenbos, L., Pissas, T., Huber, A.T., Sznitman, R., Márquez-Neila, P.: Stochastic segmentation with conditional categorical diffusion models. In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1119–1129 (2023)

2023

[26] [26]

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

Zhang, H., Luo, X., Chen, Y., Li, K.: Diffoseg: Omni medical image segmenta- tion via multi-expert collaboration diffusion model. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 128–138. Springer (2025)

2025

[27] [27]

Bioinformatics40(4), btae118 (2024)

Zhang,R.,Wu,C.,Yang,Q.,Liu,C.,Wang,Y.,Li,K.,Huang,L.,Zhou,F.:Molfes- cue: enhancing molecular property prediction in data-limited and imbalanced con- texts using few-shot and contrastive learning. Bioinformatics40(4), btae118 (2024)

2024