List of Papers Browse by Subject Areas Author List
Abstract
Ambiguous medical image segmentation often admits multiple clinically plausible delineations due to substantial inter-expert variability, and such inherent uncertainty cannot be faithfully represented by a single deterministic mask. Promptable foundation models such as SAM offer strong representations, yet under ambiguous boundaries, their predictions can be highly sensitive to prompt perturbations, and existing stochastic prompting strategies are often coarse and weakly coupled to image-side ambiguity, yielding diverse samples that may deviate from image-consistent boundaries. We propose Conditional Diffusion Prompting (CDP), which performs diffusion prompting in dense prompt embedding space prior to SAM-style decoding and conditions prompt-space denoising on stochastic image embeddings to couple prompt ambiguity with image ambiguity. Experiments on multiple ambiguous medical image segmentation benchmarks demonstrate improved distribution matching and sample quality, achieving lower GED and higher HM-IOU and Dmax. CDP provides a structured sampling mechanism for promptable decoders, enabling diverse yet image-consistent segmentations and improving practical reliability in ambiguous clinical settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5253_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/shannan-zhk/Conditional-Diffusion-Prompting
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaHon_Conditional_MICCAI2026,
author = { Zhao, Hongkai AND Gao, Jun AND Kang, Qingbo AND Lao, Qicheng},
title = { { Conditional Diffusion Prompting for Ambiguous Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposed Conditional Diffusion Prompting (CDP) to address the limitations observed in existing promptable segmentation foundation models that have the limited ability to capture complex boundary variability and produce delineations when increasing diversity. To addressed these limitations, this paper reformulates prompt generation as probabilistic distribution modeling in dense prompt embedding space and further aligns prompt uncertainty with image ambiguity by conditioning the prompt-space diffusion on stochastic image embeddings, leading to diverse yet image-consistent segmentation candidates. Experimental results conducted on the benchmarks demonstrate consistent improvements in distribution matching and sample quality and superior performance over the compared baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.While diffusion models and conditional diffusion models have been widely used and studied, the motivations behind the use of diffusion prompting and conditioning the prompt-space denoising on image embeddings make sense. The paper innovatively applied them to the ambiguous medical image segmentation task. 2.The paper is well organized and presented clearly. 3.Extensive experiments have been conducted and results demonstrates its effectiveness.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The Fig. 1 should be further improved as ‘DP’ and ‘CDP’ is not annotated in the figure. Such annotations help the reader better match the figure and description. 2.Providing source code is recommended for ensure reproducibility as the space for description of implementation is limited. 3.The compared approaches are outdated: Prob UNet(2018), PHiSeg (2019, which is actually mentioned but not compared), SSN (2020), CAR (2021), Mose (2022).
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper proposed to model prompt uncertainty through generating stochastic prompt latents and conditioned the prompt-space denosing on image embeddings to better achieved diversity and segmentation quality. The motivation makes sense and extensive experiments have conducted, including Quantitative Comparison, Qualitative Comparison and Ablation studies. However, the compared approaches are outdated, and it is inappropriate to state “outperforming state-of-the-art baselines”.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The paper proposed Conditional Diffusion Prompting (CDP) to address the limitations observed in existing promptable segmentation foundation models that have the limited ability to capture complex boundary variability and produce delineations when increasing diversity. The main concern regarding of comparison with latest SOTA methods is partly addressed. Minor typos are promised to be corrected for the camera-ready version. Overall, the paper makes a significant methodological and empirical contribution to ambiguous medical image segmentation and meets the bar for MICCAI acceptance.
Review #2
- Please describe the contribution of the paper
This paper addresses the problem of ambiguous medical image segmentation, where a single image may correspond to multiple clinically plausible masks and therefore cannot be adequately represented by a deterministic prediction. It proposes Conditional Diffusion Prompting (CDP), which models uncertainty in dense prompt embedding space and conditions prompt-space diffusion on stochastic image embeddings to generate diverse yet image-consistent segmentation samples for SAM-style decoders.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper focuses on learning a distribution of plausible masks rather than producing a single deterministic output. This is a meaningful problem formulation for clinical settings with substantial inter-expert variability. 2.The paper moves uncertainty modeling from output mask space to dense prompt embedding space. Conditioning prompt-space diffusion on stochastic image embeddings is also an intuitive way to connect prompt uncertainty with image ambiguity.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The proposed method involves a relatively long and complex pipeline, which makes the main technical contribution somewhat unclear. The final performance gain may come from several interacting components, including the semantic mask prompt generator, prompt-space diffusion, stochastic image latent modeling, or their combination, but the paper does not fully disentangle their individual contributions. As a result, it is difficult to identify a clean methodological focus and clearly understand what the primary source of improvement is. 2.While the idea of performing diffusion in dense prompt embedding space is interesting, the paper does not sufficiently explain why prompt space is a more appropriate or necessary modeling space than mask space or latent mask space. 3.The method requires DDIM sampling with 100 reverse steps, and 16 samples are generated per test image to evaluate the segmentation distribution. From a deployment perspective, the performance gain may not clearly justify this inference overhead, so the paper would benefit from additional efficiency analysis, such as per-image sampling time, memory consumption, and the trade-off between performance and sampling cost. 4.The experimental comparison lacks more analysis against closely related classes of methods. For example, compared with traditional latent-variable AMIS methods, it remains unclear whether the observed gains mainly come from the use of a promptable decoder or from the proposed diffusion prompting mechanism itself. A more structured comparison across methodological categories would make the empirical claims more convincing.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Although the paper addresses an important problem and presents an interesting idea of modeling ambiguity in prompt embedding space, the methodological contribution is not yet sufficiently well organized. In particular, the pipeline contains several interacting components, making it unclear whether the gains mainly come from diffusion prompting itself or from other auxiliary designs, and the paper does not fully justify why prompt space is a more appropriate modeling space than mask space. In addition, the experimental analysis is still incomplete, especially in terms of efficiency and more comparisons with closely related classes of ambiguous segmentation methods.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal sufficiently addresses my main concerns. The authors provide concrete efficiency analysis, including runtime, memory, FLOPs, and sampling-cost trade-offs, which clarifies the practical overhead of the proposed method. They also strengthen the comparison with additional related methods such as ATFM and plan to include PHiSeg, making the empirical evaluation more complete. The explanation that dense prompt embeddings are the native control variable of SAM-style decoders also makes the motivation for prompt-space diffusion clearer. Although I still think the paper would benefit from more detailed disentanglement of the semantic prompt generator, stochastic image latent modeling, and diffusion prompting components, the Base/DP/CDP ablation and the rebuttal clarification reasonably support the contribution of the proposed diffusion prompting framework. Given the importance of the problem, the interesting methodological direction, and the consistent empirical improvements, I update my final recommendation to Accept.
Review #3
- Please describe the contribution of the paper
The paper introduces Conditional Diffusion Prompting (CDP), a framework for modeling ambiguity in medical image segmentation by performing diffusion-based denoising in the dense prompt embedding space of SAM-style models. Unlike prior approaches that model uncertainty in mask space or inject unstructured noise into prompts, the method reformulates prompt generation as a probabilistic latent modeling problem. It first proposes Diffusion Prompting (DP) to learn a structured distribution over prompt embeddings via iterative denoising, and then extends this to CDP by conditioning the diffusion process on stochastic image latents, explicitly coupling prompt uncertainty with image-side ambiguity. This ensures that generated segmentation samples are both diverse and image-consistent, effectively capturing the one-to-many nature of expert annotations. Across multiple medical benchmarks, CDP achieves state-of-the-art performance in distribution matching and sample fidelity, outperforming both probabilistic segmentation models and stochastic prompt-based baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.This paper introduces diffusion in dense prompt embedding space, offering a structured alternative to mask-space modeling or random prompt perturbations for varied medical imaging tasks. 2.CDP conditions prompt diffusion on stochastic image latents, ensuring generated segmentations are both diverse and aligned with image ambiguity, addressing a key limitation of prior SAM-based methods. 3.Achieves competitive GED, HM-IoU and Dmax on multiple benchmarks (LIDC, ISIC, Prostate, Pancreatic-lesion), demonstrating improved distribution matching and fidelity. Ablation studies show consistent gains from DP and CDP components, and sensitivity analyses indicate robustness to hyperparameters and sampling count.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The paper does not clearly establish why diffusion in prompt space is superior to mask-space diffusion or other latent modeling approaches. There should at least be some discussion around that if not experimental validation. 2.This relies on a frozen mask generator and SAM-style encoder, which may propagate biases or limit adaptability if the initial prompt distribution is suboptimal. What should be the necessary steps for that ? Is that a major issue? Expanding on this, while GED and Dmax are reported, the work lacks explicit quantitative analysis of mode diversity or collapse. The authors mention this qualitatively in Section 4.2 and while Dmax shows the best-case scenario, it doesn’t prove diversity. The paper would greatly benefit from an analysis using metrics such as Self-Similarity or Bhattacharyya Distance, making clear how distinct the generated samples truly are in a quantitative sense. 3.Diffusion-based sampling introduces significant inference overhead, but the paper does not provide runtime or FLOPs comparisons with baselines. 4.Several important implementation details required for reproducibility for example, specific architecture of the lightweight denoiser, the noise schedule parameters for the diffusion process and the dimensionality of the latent variables and many other hyperparameters as well as design choices are missing from the “Implementation Details” section.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents an approach by modeling ambiguity in dense prompt embedding space using diffusion, and further coupling it with image uncertainty via CDP. It demonstrates consistent empirical improvements across multiple datasets and metrics, supported by ablation and sensitivity analyses. However, the paper does not clearly justify why prompt-space diffusion is preferable to alternative approaches, and lacks explicit quantitative evaluation of diversity quality. Additionally, efficiency comparisons and key implementation details for reproducibility are missing. Overall, these factors place the paper slightly above the acceptance threshold, leading to a Weak Accept recommendation prior to rebuttal.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have answered my concerns including clarification of the motivation, discussion of limitations, diversity analysis, and computational costs. They have said that the source code will be released for reproducibility in the future upon acceptance
Author Feedback
We sincerely thank the AC and reviewers for the constructive comments and address the raised concerns below.
1.Comparisons (R1, R2) We agree that more recent and structurally related methods should be better covered. In the manuscript, we include comparisons with both traditional latent-variable AMIS methods (Table 1) and A-SAM [2] (NeurIPS 2024), a recent promptable AMIS baseline. On LIDC, our method reduces GED by 9.8% and improves HM-IoU from 0.768 to 0.839 (+9.2%) compared with A-SAM, indicating that the improvement is not simply due to adopting a promptable decoder. We further compare with ATFM [27] (AAAI 2026), a more recent flow-matching-based mask-space generative method, where our method still achieves superior HM-IoU performance (0.839 vs. 0.667), and improves best-match Dice by approximately 3.0%. Finally, both A-SAM and ATFM have previously demonstrated advantages over PHiSeg (2019). We will also include PHiSeg in the revision for a more comprehensive comparison.
2.Why Diffusion in Prompt Space (R2, R3) Prompt space is particularly suitable for SAM-style decoders because it serves as the decoder’s native control variable, unlike mask-space modeling performed after decoding. Modeling this space preserves the promptable decoder prior while allowing stochastic sampling. Previous work, such as A-SAM [2], explicitly models prompt-space ambiguity and demonstrates superior performance. Building on this, our method further replaces Gaussian prompt sampling with diffusion denoising in dense prompt space, conditioned on stochastic image embeddings, such that prompt uncertainty is aligned with underlying image ambiguity.
3.Inference Overhead and Efficiency Analysis (R2, R3) We thank the reviewer for the valuable suggestion. We provide the runtime, memory, sampling-cost trade-off, and FLOPs comparison with A-SAM. Our method uses higher computation than A-SAM, with 49.54 vs. 22.66 GFLOPs/sample. However, this overhead is tunable: the 100-step DDIM setting is our upper-bound setting for stronger distribution modeling. Reducing it to 10 steps lowers the runtime from 3.464s to 0.721s with 16 samples/image, while peak memory remains about 9.42 GB; with 100 steps, the runtime scales from 0.220s for 1 sample to 3.464s for 16 samples. Together with Table 3, these results show a clear cost-quality trade-off: fewer DDIM steps or samples substantially reduce latency, while the full setting provides stronger distribution modeling when higher fidelity is desired. We will include these efficiency analyses and discussions in the revision.
4.Reproducibility and Source Code (R1, R3) We will specify the key settings (e.g., ViT-B SAM, LoRA rank 4, 256-d latent variables) and release the code and models upon acceptance to ensure reproducibility.
5.Fig. 1 Annotation (R1) We thank the reviewer for the suggestion. We will explicitly annotate “DP” and “CDP” in Fig. 1 to match the description.
6.Ablation Clarity (R2) We apologize for the insufficient clarity of the ablation analysis in the current manuscript. In Table 2, we disentangle the contributions of each major component, including the semantic mask prompt generator (denoted as “Base”, 0.7823 HM-IoU), prompt-space diffusion (denoted as “DP”, 0.8363 HM-IoU), and stochastic image latent modeling (denoted as “CDP”, 0.8493 HM-IoU). These results show that the primary performance improvement does not originate from the prompt generator alone, but mainly from the proposed diffusion prompting framework. To further improve clarity, we will revise the manuscript to more explicitly emphasize the role and contribution of each component.
7.Diversity Analysis (R3) We provide Self-Similarity and Bhattacharyya Distance on LIDC to quantify mode diversity/collapse beyond GED and Dmax. Compared with the baseline, CDP reduces Self-Similarity from 0.469 to 0.387 and increases Bhattacharyya Distance from 1.301 to 1.525, indicating lower sample similarity and larger distributional separation.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This work received mixed reviews, with key concerns focusing on the experimental comparisons and the diversity of evaluation results. In particular, recent methods are not included in the comparisons, and clearer ablation studies are needed to highlight the effectiveness of the core diffusion prompting. In addition, reproducibility, computational costs, and theoretical analysis require further clarification. Therefore, the authors are invited to submit a rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have sufficiently addressed the reviewers’ comments in the rebuttal, and all reviewers recommended acceptance. The main contribution of this work lies in the design of conditional prompts within the diffusion process for ambiguous medical image segmentation. The performance is good, and the authors are required to further improve the manuscript in the camera-ready version. Therefore, a final acceptance recommendation is given.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
All three reviewers indicate that most of their concerns have been properly addressed and proposed to accept the paper. Moreover, they mention an interesting methdological contribution and consistent empirical improvements on an important problem. While still several directions of further improving the work remain open (e.g. R2: detailed disentanglement of the semantic prompt generator, stochastic image latent modeling, and diffusion prompting components, the Base/DP/CDP ablation), all reviewers agree that the work meets the bar for MICCAI after the rebuttal phase.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The rebuttal clarified major concerns raised by reviewers, and all reviewers agree on its acceptance. This guarantees a clear accept.