From Variability To Accuracy: Conditional Bernoulli Diffusion Models with Consensus-Driven Correction for Thin Structure Segmentation

Helen Hong; Jinseo An; Kyu Won Shim; Min Jin Lee

REVIEW 4 major objections 7 minor 16 references

From Variability To Accuracy: Conditional Bernoulli Diffusion Models with Consensus-Driven Correction for Thin Structure Segmentation

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that consensus among 200 conditional Bernoulli diffusion segmentations, corrected by an energy coupling position, consensus, and gradient direction, lifts orbital medial wall recall from 84.60% to 87.83% while holding…

desk verdict The recall gain is a real but likely precision/recall tradeoff, and the only genuinely new piece is the consensus-driven correction energy; the diffusion part is BerDiff and multi-annotation training is not new. read the letter →

arxiv 2507.12985 v1 pith:TPP7HKAL submitted 2025-07-17 eess.IV cs.CV

Jinseo An , Min Jin Lee , Kyu Won Shim , Helen Hong This is my paper

classification eess.IVcs.CV

keywords orbitalbonesegmentationBernoullidiffusionmodelconsensus-drivencorrectioninter-observervariabilitythinstructureuncertaintymapfacialCT

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 claims that under-segmentation of thin, ambiguous orbital bone structures in facial CT can be repaired by exploiting disagreement among many plausible segmentations generated by a conditional Bernoulli diffusion model, rather than by relying on a single deterministic prediction. It trains the diffusion model on three expert annotations per image so the model learns inter-observer variability, then aggregates 200 sampled segmentations into a consensus-driven uncertainty map. Pixels where the samples disagree are reclassified by minimizing an energy that couples position proximity, consensus-level similarity, and gradient-direction similarity. The reported result is that this correction raises recall on the orbital medial wall from 84.60% to 87.83% against the BerDiff baseline while keeping Dice at 86.31%, and improves continuity of thin structures. If true, this would automate a manual correction step and make diffusion-based segmentation more usable for surgical planning and custom implant design.

What carries the argument

The load-bearing machinery is a two-stage pipeline. First, a conditional Bernoulli diffusion model, trained on three different expert annotations per CT image with one annotation randomly selected each iteration, generates 200 diverse plausible segmentations; the forward process adds Bernoulli rather than Gaussian noise, matching the binary nature of masks. Second, the 200 outputs are averaged into a consensus-driven uncertainty map, and an energy function $E(x)=\sum_i \psi_i(x_i)+\sum_{ij}\psi_{ij}(x_i,x_j)$ corrects low-consensus pixels: the unary term is $-\log P(x_i)$ derived from the uncertainty map, and the pairwise term blends three Gaussian kernels over position proximity $p$, consensus level $c$, and gradient direction $d$, with parameters set empirically. This energy propagates confident labels into uncertain thin regions while respecting directional continuity, which is what rescues the orbital medial wall.

What would settle it

Have new experts, not involved in the original three annotations, independently re-annotate the 14 test CTs, fuse their masks with STAPLE, and re-run the BerDiff versus corrected comparison on the orbital medial wall; if the corrected recall no longer exceeds BerDiff's 84.60% by a significant margin, the correction is tuned to the training-time reference rather than to real boundaries.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the variability a Bernoulli diffusion model produces across repeated samples is signal, not noise: low-consensus pixels mark exactly the ambiguous thin-bone regions where deterministic CNNs and single diffusion outputs under-segment. The authors show that training the diffusion model on multiple annotation patterns per CT image lets it generate a family of plausible masks, and that a consensus-driven correction step, which re-labels low-confidence pixels using position proximity, consensus-level similarity, and gradient-direction similarity, repairs disconnected medial-wall and orbital-floor structures. Quantitatively, the corrected output beats CNN baselines on all metrics and raises medial-wall recall from 84.60% (BerDiff) to 87.83% (ours), with p<0.001, while keeping Dice at 86.31%; orbital-floor recall rises from 92.15% to 93.24%. The intended consequence is a segmentation pipeline that preserves thin structures and automates what was previously manual correction.

Load-bearing premise

The load-bearing premise is that the reference standard produced by the STAPLE algorithm, which merges the same three annotator masks used to train the diffusion model, is an unbiased ground truth; if that merged reference mostly reflects the majority or one annotator's style, the reported recall gains may show agreement with that reference rather than with true anatomical boundaries.

Editorial extensions

If this is right

If the central claim is right, diffusion ensembles become a correction mechanism rather than just a way to sample plausible masks: disagreement localizes the ambiguous regions that need fixing.
The method reduces manual correction effort in orbital implant planning because the consensus-driven step automates re-labeling of under-segmented thin bone.
The statistically significant recall gain on the orbital medial wall, from 84.60% to 87.83% with p<0.001 and Dice preserved at 86.31%, suggests the approach specifically attacks under-segmentation rather than merely improving overall overlap.
Because the diffusion model is trained on multiple annotator masks, the framework converts inter-observer variability from a nuisance into training signal for generating plausible variation.
The same pipeline is claimed to apply to image-guided surgical planning and surgery for orbital fracture reconstruction.

Reading between the lines

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

Beyond the paper, the consensus-correction principle may transfer to other thin or ambiguous structures, such as the cribriform plate, orbital fissures, or small vessels, whenever the boundary has a consistent local direction; this could be tested by running the same pipeline unchanged on those structures.
Beyond the paper, the reported hyperparameters ($\theta_\alpha=80$, $\theta_\beta=60$, $\theta_\gamma=2$, $\theta_\delta=3$, $w_1=15$, $w_2=1$) were chosen from experiments, so the correction gain would be more convincing if it persisted across a sweep of these values or on an independent validation set.
Beyond the paper, because the correction energy assumes low consensus means error, a stress test would be to inject synthetic disagreements known to be correct; if correction still flips those pixels, the assumption is false.
Beyond the paper, the 200-sample ensemble suggests a cost-quality frontier: reporting performance versus the number of diffusion samples would tell clinicians whether the improvement is affordable under real time constraints.

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

Two things to know before reading: the medial-wall recall gain is real but looks like a precision/recall tradeoff, not an accuracy win; and the genuinely new part is the correction energy, not the diffusion model. The diffusion backbone is BerDiff, and training on multiple annotations per image is Rahman et al.'s approach. What's new is the consensus-driven correction that combines position, consensus, and gradient direction in a CRF-style energy, plus the orbital bone application.

The paper does some things well. The energy formulation is reasonable, the idea of using the ensemble's uncertainty map to drive correction is sensible, and the qualitative continuity improvement in thin structures is plausible. The table is honest enough to let you see the tradeoff yourself: on the medial wall, recall goes from 84.60 to 87.83 while precision drops from 88.16 to 85.38, and DSC moves only 0.17. On the floor, DSC actually decreases. The abstract says the method 'outperforms existing methods,' but that overstates things when BerDiff has higher floor DSC and higher precision on both ROIs.

The main soft spot is the reference standard. STAPLE is built from the same three annotator masks used to train the diffusion model, so the correction is effectively rewarded for moving toward the consensus of those same three raters. That doesn't invalidate the comparison against baselines, all measured on the same yardstick, but it does mean 'accuracy' here really means 'agreement with training annotators.' An independent reference, or at least a held-out annotator, would make the recall claim much stronger.

Other issues: hyperparameters for the correction energy are chosen 'based on experimental results' with no validation split, on only 14 test patients. There's no ablation isolating the three pairwise terms, no comparison against Rahman et al. (the closest ambiguity-learning baseline), and no code or data. Individually these are minor to moderate; together they undercut the strength of the claims.

This paper is for someone working on thin-structure segmentation in CT or on ambiguity-aware diffusion segmentation. It's a competent incremental paper that deserves a serious referee, but it needs more rigorous evaluation before acceptance. My recommendation: engage with it, but ask for a validation strategy, an ablation, and a more careful reading of the tradeoff.

Referee Report

4 major / 7 minor

Summary. The paper proposes a segmentation framework for thin orbital bone structures in CT. A conditional Bernoulli diffusion model is trained on three annotator masks per image, then at inference 200 segmentations are sampled and averaged into a consensus/uncertainty map. A post-processing energy minimization (unary term from the consensus map, pairwise terms encoding position proximity, consensus similarity, and gradient direction similarity) corrects low-consensus pixels. Experiments on 71 facial CTs (57 train/14 test) report DSC, recall, and precision for the orbital medial wall and floor, with the main claimed gain being significantly higher recall on the medial wall (87.83 vs 84.60 for BerDiff) while preserving DSC.

Significance. If the reported effect is robust, the idea of using multi-sample diffusion consensus as a spatially varying prior for a corrective energy is a reasonable and potentially useful contribution to segmentation of anatomies with high inter-observer variability. The paper's strengths are that it directly addresses a clinically relevant problem, uses a Bernoulli diffusion formulation appropriate for binary masks, and explicitly targets thin-structure continuity. However, the current experimental design cannot yet separate a genuine accuracy gain from a rater-bias artifact: the reference standard is built from the same three annotator masks used to train the model, the correction hyperparameters are selected without a validation split, and the test set has only 14 patients. The reported precision decreases in both ROIs and the floor DSC decreases relative to BerDiff, so the 'outperforms' claim is not established.

major comments (4)

[Section 3.2 / Table 1] The evaluation reference is a STAPLE fusion of the same three annotator masks that are used to train the diffusion model. Because the model is trained to reproduce those annotators' segmentation distribution, its consensus map is not an independent source of anatomical truth, and correcting low-consensus pixels toward the ensemble consensus may simply move predictions toward the STAPLE majority. Please evaluate with a held-out annotator (e.g., train on two annotators and test against the third), an independent expert re-annotation, or at least a sensitivity analysis showing that the medial-wall recall gain persists under different reference constructions.
[Section 2.2, Eq. (3)] The correction energy has six hyperparameters (θα, θβ, θγ, θδ, w1, w2) that are set 'based on experimental results' with no validation split or ablation. On a 14-patient test set, this creates a risk that the reported recall improvement is an artifact of tuning on the test set. Please add a validation-based selection procedure or an ablation/sensitivity table, and report the chosen values' stability.
[Table 1] The conclusion that the method 'outperforms other comparison methods' is not supported across all metrics. Orbital floor DSC is lower than BerDiff (91.36 vs 91.53), precision is lower in both ROIs (85.38 vs 88.16 on medial wall; 90.08 vs 91.37 on floor), and only medial-wall recall is marked as statistically significant. Please report paired statistical tests for every metric, with confidence intervals, and discuss whether the recall gain justifies the precision loss for the intended surgical-planning use case.
[Section 3.3 / Fig. 3] The qualitative claim that the method 'corrects challenging regions' relies on the consensus map generated by the model itself. Without evidence that low-consensus pixels correspond to actual segmentation errors (e.g., by checking against CT intensity boundaries or an independent reference), the correction may be reinforcing the model's own biases. Please add a quantitative analysis of the correlation between low consensus and true error, or an oracle experiment.

minor comments (7)

[Title] The title contains a typo: 'T o' should be 'To'.
[Figure 1] The subscript glyphs (e.g., 'g1840', 'g1839') are corrupted; please fix the figure so that all symbols render correctly.
[Section 2.2] Define c_i and d_i explicitly, and specify the image from which gradients are computed (CT intensity vs. consensus map).
[Section 3.2] Clarify how the 200 segmentations are combined into 'Ours': is it a single corrected mask from the mean probability map, and how is the energy minimized (graph cut, ICM, etc.)?
[Section 3.2] Add a baseline that simply averages the 200 samples with a fixed threshold, to isolate the contribution of the correction step.
[Conclusion] The statement 'improvements of up to 4.03% in DSC and 8.13% in recall' refers to CNN baselines; please state the comparison point openly to avoid overstating the improvement over the best diffusion baseline.
[References] References [2] and [8] are conference presentations without full details; please provide accessible versions or more information about the study design and results.

Circularity Check

0 steps flagged · score 2.0 of 10

No derivation in the paper reduces to its own inputs by construction; the score reflects a non-circular but non-trivial coupling between the training annotations and the STAPLE reference, not a circularity finding.

full rationale

The method's chain is: conditional Bernoulli diffusion trained on three annotations per CT (Sec. 3.1), 200 sampled segmentations aggregated into a consensus probability map (Sec. 2.1, Fig. 2), and a CRF-style energy minimization whose unary term is -log P(x_i) and pairwise terms use position, consensus, and gradient (Sec. 2.2, Eqs. 1-3). Each stage is explicitly stated as using the model's own uncertainty, so using the consensus map as correction input is a design choice, not a hidden reuse of the target. No equation in the paper has the form 'predicted quantity = fitted parameter' and no fitted value is renamed a prediction. The hyperparameters (theta_alpha=80, theta_beta=60, theta_gamma=2, theta_delta=3, w1=15, w2=1) are said to be set 'based on experimental results' (Sec. 2.2), which is an overfitting/validation concern, but it does not make the reported DSC/recall values true by construction. The self-citations (refs [2,3]) motivate the problem and supply a comparison baseline; they do not carry the load of the correction claim, and no uniqueness theorem is imported from the authors' prior work. The one significant non-circular weakness is that Section 3.2 builds the STAPLE reference from the same three annotator masks used for training the diffusion model, so the evaluation reference is not independent of the training distribution; this can bias measured recall/precision and should be weighed as a benchmark-validity risk, not as mathematical circularity. Overall, the derivation is self-contained and no circular step meets the evidentiary bar of Eq. X = Eq. Y by construction.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim relies on the assumptions that annotator variability is learnable, that uncertainty correlates with error, and that the STAPLE reference is unbiased. The only fitted quantities are the CRF kernel hyperparameters and the ensemble size; the diffusion backbone is taken from BerDiff. No new physical entities are introduced.

free parameters (2)

Correction energy weights and kernel bandwidths = w1=15, w2=1, θα=80, θβ=60, θγ=2, θδ=3
Set based on experimental results (Section 2.2); they directly control how strongly consensus, position, and gradient features drive correction, so the reported accuracy depends on these hand-tuned values.
Number of diffusion samples per image = 200
The model is run 200 times to form the consensus map; no sensitivity analysis is provided, so the stability of the uncertainty map with this choice is untested.

assumptions (5)

domain assumption Multiple manual annotations per image sample the true distribution of plausible segmentations in ambiguous regions
The method depends on this to train the diffusion model and to produce a meaningful consensus map; if the three annotators do not span the real variability, the model cannot learn it.
domain assumption Low consensus across diffusion outputs indicates segmentation errors that correction should fix
The consensus-driven correction targets low-consensus pixels; if uncertainty is not correlated with anatomical error, correction may introduce new errors.
domain assumption STAPLE-combined reference from the same three annotators is a valid ground truth
All metrics are computed against the STAPLE reference, so the evaluation inherits any bias in STAPLE toward majority annotation patterns.
ad hoc to paper Gradient direction similarity helps preserve thin bone structures
The gradient term in Eq. 3 is justified qualitatively by thin-structure directionality, without quantitative evidence that it helps.
standard math Gaussian kernel smoothness and Potts compatibility functions behave as in dense CRFs
The energy function in Eq. 1-3 relies on standard CRF assumptions that similar nearby pixels should share labels; the paper does not verify these assumptions for CT bone boundaries.

how reviews work

0 comments

Cite this review

Pith. "Pith review of From Variability To Accuracy: Conditional Bernoulli Diffusion Models with Consensus-Driven Correction for Thin Structure Segmentation." pith.science (2026). https://pith.science/paper/TPP7HKAL

@misc{pith2026250712985,
  author       = {Pith},
  title        = {Pith review of: From Variability To Accuracy: Conditional Bernoulli Diffusion Models with Consensus-Driven Correction for Thin Structure Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TPP7HKAL}},
  note         = {Machine review of arXiv:2507.12985}
}

read the original abstract

Accurate segmentation of orbital bones in facial computed tomography (CT) images is essential for the creation of customized implants for reconstruction of defected orbital bones, particularly challenging due to the ambiguous boundaries and thin structures such as the orbital medial wall and orbital floor. In these ambiguous regions, existing segmentation approaches often output disconnected or under-segmented results. We propose a novel framework that corrects segmentation results by leveraging consensus from multiple diffusion model outputs. Our approach employs a conditional Bernoulli diffusion model trained on diverse annotation patterns per image to generate multiple plausible segmentations, followed by a consensus-driven correction that incorporates position proximity, consensus level, and gradient direction similarity to correct challenging regions. Experimental results demonstrate that our method outperforms existing methods, significantly improving recall in ambiguous regions while preserving the continuity of thin structures. Furthermore, our method automates the manual process of segmentation result correction and can be applied to image-guided surgical planning and surgery.

Figures

Figures reproduced from arXiv: 2507.12985 by the authors.

**Figure 1.** Illustration of our conditional Bernoulli diffusion model adds random noise through Bernoulli noise sampling. In the reverse process, the model progressively removes noise to generate segmentations, conditioned on the corresponding CT images. The CT images used as conditions provide intensity and context information, enabling the model to understand anatomical structures during the denoising process. An advantage of… view at source ↗

**Figure 2.** Process of generating a consensus-driven uncertainty map uncertainty map, which provides valuable insight into potentially correct segmentation areas. Consensus-driven correction minimizes the energy function in Eq. 1, which combines the unary potential (Eq. 2) and pairwise potential (Eq. 3). The unary potential is derived from the consensus level from the map, providing a pixelwise measure of segmentation confiden… view at source ↗

**Figure 3.** Representative orbital bone segmentation results. Consensus-driven uncertainty map is generated by accumulating multiple segmentations using a conditional Bernoulli diffusion model (BerDiff). ‘BerDiff’ in 4th column is the mean of multiple segmentations. The red box is zoomed in to emphasize the thin region within the white box. Acknowledgments. This research was partially supported by a grant of the Korea Health T… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 12 canonical work pages

[1]

arXiv prepr int (2021), https://arxiv.org/abs/2112.00390

Amit, T., Shaharbany, T., Nachmani, E., Wolf, L.: Segdiﬀ: Image seg- mentation with diﬀusion probabilistic models. arXiv prepr int (2021), https://arxiv.org/abs/2112.00390

arXiv 2021
[2]

Presented at Radiological Society of North America (RSNA) (2023)

An, J., Lee, M.J., Hong, H., Shim, K.W.: Evaluation of segm entation performance using multiple reference standards for accurate orbital bo ne modeling in cranio- maxillofacial surgery: based on msda-net deep learning seg mentation. Presented at Radiological Society of North America (RSNA) (2023)

work page 2023
[3]

Journal of the Korea Computer Graphics Society 29(2), 13–20 (2023)

An, J., Lee, M.J., Shim, K.W., Hong, H.: Orbital bone segme ntation using improved skip connection of u-net structure in facial ct images. Journal of the Korea Computer Graphics Society 29(2), 13–20 (2023). https://doi.org/10.15701/kcgs.2023.29.2.13

work page doi:10.15701/kcgs.2023.29.2.13 2023
[4]

In: Medical Image Com puting 8 J

Chen, T., Wang, C., Shan, H.: Berdiﬀ: Conditional bernoul li diﬀusion model for medical image segmentation. In: Medical Image Com puting 8 J. An et al. and Computer Assisted Intervention – MICCAI 2023. Lecture N otes in Computer Science, vol. 14223, pp. 491–501. Springer, Cha m (2023). https://doi.org/10.1007/978-3-031-43901-8_47

work page doi:10.1007/978-3-031-43901-8_47 2023
[5]

IEEE Transactions on Medic al Imaging 43(10), 3570–3583 (2024)

Chen, T., Wang, C., Chen, Z., Lei, Y., Shan, H.: Hidiﬀ: Hybr id diﬀusion framework for medical image segmentation. IEEE Transactions on Medic al Imaging 43(10), 3570–3583 (2024). https://doi.org/10.1109/TMI.2024.3424471

arXiv 2024
[6]

Craniomaxillofa cial trauma & reconstruc- tion 6(3), 161–169 (2013)

Cho, R.I., Davies, B.W.: Combined orbital ﬂoor and medial wall fractures in- volving the inferomedial strut: repair technique and case s eries using preshaped porous polyethylene/titanium implants. Craniomaxillofa cial trauma & reconstruc- tion 6(3), 161–169 (2013). https://doi.org/10.1055/s-0033-1343785

work page doi:10.1055/s-0033-1343785 2013
[7]

In: Medical Image Computing and Computer Assisted Int ervention – MIC- CAI 2023

Chowdary, G.J., Yin, Z.: Diﬀusion transformer u-net for m edical image segmenta- tion. In: Medical Image Computing and Computer Assisted Int ervention – MIC- CAI 2023. Lecture Notes in Computer Science, vol. 14223. Spr inger, Cham (2023). https://doi.org/10.1007/978-3-031-43901-8_59

work page doi:10.1007/978-3-031-43901-8_59 2023
[8]

Presented a t The Korean Society of 3D Printing in Medicine (2022)

Kim, H., et al.: Comparative evaluation of inter-observe r variability in manual seg- mentation of orbital bone from facial ct images. Presented a t The Korean Society of 3D Printing in Medicine (2022)

work page 2022

Show all 16 references

[9]

Journal of Cranio-Maxillofacial Surgery 47(6), 959–967 (2019)

Kim, H., et al.: Three-dimensional orbital wall modeling using paranasal sinus segmentation. Journal of Cranio-Maxillofacial Surgery 47(6), 959–967 (2019). https://doi.org/10.1016/j.jcms.2019.03.028

2019 doi
[10]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVP R)

Rahman, A., Valanarasu, J.M.J., Hacihaliloglu, I., Pat el, V.M.: Ambiguous medical image segmentation using diﬀusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVP R). pp. 11536– 11546 (2023)

2023
[11]

In: Medical Image Computing and Com puter-Assisted In- tervention – MICCAI 2015

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolut ional networks for biomedi- cal image segmentation. In: Medical Image Computing and Com puter-Assisted In- tervention – MICCAI 2015. Lecture Notes in Computer Science , vol. 9351, pp. 234–

2015
[12]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Reco gnition (CVPR)

Wang, Z., Wang, J., Liu, Z., Qiu, Q.: Binary latent diﬀusi on. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Reco gnition (CVPR). pp. 22576–22585 (2023)

2023
[13]

IEEE Transactions on Medical Imaging 23(7), 903–921 (2004)

Warﬁeld, S.K., Zou, K.H., Wells, W.M.: Simultaneous tru th and perfor- mance level estimation (staple): an algorithm for the valid ation of image seg- mentation. IEEE Transactions on Medical Imaging 23(7), 903–921 (2004). https://doi.org/10.1109/TMI.2004.828354

2004
[14]

Proceedings of the AAAI Conference on Artiﬁcial Intelligence 38(6), 6030–6038 (2024)

Wu, J., Ji, W., Fu, H., Xu, M., Jin, Y., Xu, Y.: Medsegdiﬀ-v 2: Diﬀusion-based medical image segmentation with transform er. Proceedings of the AAAI Conference on Artiﬁcial Intelligence 38(6), 6030–6038 (2024). https://doi.org/10.1609/aaai.v38i6.28418

2024 doi
[15]

Knowledge-Based Systems 301, 112350 (2024)

Yan, P., et al.: Cold segdiﬀusion: A novel diﬀusion model for med- ical image segmentation. Knowledge-Based Systems 301, 112350 (2024). https://doi.org/10.1016/j.knosys.2024.112350

2024
[241]

https://doi.org/10.1007/978-3-319-24574-4_28

Springer, Cham (2015). https://doi.org/10.1007/978-3-319-24574-4_28

2015 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

arXiv prepr int (2021), https://arxiv.org/abs/2112.00390

Amit, T., Shaharbany, T., Nachmani, E., Wolf, L.: Segdiﬀ: Image seg- mentation with diﬀusion probabilistic models. arXiv prepr int (2021), https://arxiv.org/abs/2112.00390

arXiv 2021

[2] [2]

Presented at Radiological Society of North America (RSNA) (2023)

An, J., Lee, M.J., Hong, H., Shim, K.W.: Evaluation of segm entation performance using multiple reference standards for accurate orbital bo ne modeling in cranio- maxillofacial surgery: based on msda-net deep learning seg mentation. Presented at Radiological Society of North America (RSNA) (2023)

work page 2023

[3] [3]

Journal of the Korea Computer Graphics Society 29(2), 13–20 (2023)

An, J., Lee, M.J., Shim, K.W., Hong, H.: Orbital bone segme ntation using improved skip connection of u-net structure in facial ct images. Journal of the Korea Computer Graphics Society 29(2), 13–20 (2023). https://doi.org/10.15701/kcgs.2023.29.2.13

work page doi:10.15701/kcgs.2023.29.2.13 2023

[4] [4]

In: Medical Image Com puting 8 J

Chen, T., Wang, C., Shan, H.: Berdiﬀ: Conditional bernoul li diﬀusion model for medical image segmentation. In: Medical Image Com puting 8 J. An et al. and Computer Assisted Intervention – MICCAI 2023. Lecture N otes in Computer Science, vol. 14223, pp. 491–501. Springer, Cha m (2023). https://doi.org/10.1007/978-3-031-43901-8_47

work page doi:10.1007/978-3-031-43901-8_47 2023

[5] [5]

IEEE Transactions on Medic al Imaging 43(10), 3570–3583 (2024)

Chen, T., Wang, C., Chen, Z., Lei, Y., Shan, H.: Hidiﬀ: Hybr id diﬀusion framework for medical image segmentation. IEEE Transactions on Medic al Imaging 43(10), 3570–3583 (2024). https://doi.org/10.1109/TMI.2024.3424471

arXiv 2024

[6] [6]

Craniomaxillofa cial trauma & reconstruc- tion 6(3), 161–169 (2013)

Cho, R.I., Davies, B.W.: Combined orbital ﬂoor and medial wall fractures in- volving the inferomedial strut: repair technique and case s eries using preshaped porous polyethylene/titanium implants. Craniomaxillofa cial trauma & reconstruc- tion 6(3), 161–169 (2013). https://doi.org/10.1055/s-0033-1343785

work page doi:10.1055/s-0033-1343785 2013

[7] [7]

In: Medical Image Computing and Computer Assisted Int ervention – MIC- CAI 2023

Chowdary, G.J., Yin, Z.: Diﬀusion transformer u-net for m edical image segmenta- tion. In: Medical Image Computing and Computer Assisted Int ervention – MIC- CAI 2023. Lecture Notes in Computer Science, vol. 14223. Spr inger, Cham (2023). https://doi.org/10.1007/978-3-031-43901-8_59

work page doi:10.1007/978-3-031-43901-8_59 2023

[8] [8]

Presented a t The Korean Society of 3D Printing in Medicine (2022)

Kim, H., et al.: Comparative evaluation of inter-observe r variability in manual seg- mentation of orbital bone from facial ct images. Presented a t The Korean Society of 3D Printing in Medicine (2022)

work page 2022