List of Papers Browse by Subject Areas Author List
Abstract
Recently, diffusion models have shown promising performance in medical image anomaly detection. However, most existing methods insufficiently model complex anatomical structures and irregular lesion morphologies. To address this limitation, we propose a \textit{Direction-Guided Shape-Scale Aware Diffusion Model} for medical anomaly detection. The proposed model introduces a \textit{Direction-Guided Multi-Scale} residual module to decouple directional features and aggregate multi-scale receptive fields, thereby enhancing anatomical shape representation. In addition, an \textit{Anomaly-Sensitive Skip Fusion} strategy is developed to guide feature fusion using high-frequency structural priors, enabling the network to better focus on fine-grained lesion boundaries and abnormal details. Furthermore, a lightweight architecture is further adopted to reduce the computational burden of iterative diffusion inference. Experimental results demonstrate that the proposed method achieves superior detection accuracy and inference efficiency compared with existing approaches. Our code is available at https://github.com/guyxlouspg/AnoSSPD.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1676_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/guyxlouspg/AnoSSPD
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZheShe_ShapeScale_MICCAI2026,
author = { Zheng, Shenhai AND Tan, Dongtao AND Li, Weisheng AND Li, Laquan},
title = { { Shape-Scale Aware Diffusion Models for Weakly Supervised Medical Image Anomaly Detection } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes AnoSSPD, a shape- and scale-aware lightweight diffusion model for medical anomaly detection. Its primary contributions include: 1.DMS Module: Enhancing the modeling of complex lesion morphology and scales through a Direction-guided Multi-Scale module. 2.ASSF Mechanism: Designing an Anomaly-Sensitive Skip Fusion mechanism to improve detection precision, particularly at the boundaries of anomalous regions. 3.Efficiency: Constructing a lightweight architecture based on PDS-Conv, which reduces computational overhead while maintaining high performance.
- 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.
- The proposed Direction-guided Multi-Scale (DMS) module enhances modeling of lesion shape and scale via directional decomposition and multi-scale aggregation. This offers a clear innovation over traditional isotropic U-Net convolutions, better capturing the irregular and directional nature of medical lesions.
- Addressing the high inference costs of diffusion models, this work introduces a lightweight PDS-Conv backbone. The reported reduction in complexity without sacrificing performance ensures the method is viable for resource-constrained, real-world medical deployments.
- The evaluation is robust, spanning BraTS 2021, BraTS-PEDs 2024, and ATLAS v2.0.By utilizing diverse metrics (Dice, IoU, AUPRC) across Mixed and Unhealthy settings, the results are comprehensive. Systematic ablation studies further confirm the effectiveness of the DMS module, ASSF, and the lightweight architecture.
- 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.
- Despite strong results across multiple datasets, the analytical depth remains somewhat limited. For instance, while the paper claims the DMS module improves directional and multi-scale modeling, it lacks a detailed analysis of which specific lesion types, anatomical structures, or boundary complexities yield the greatest performance gains.
- The innovation of this work lies primarily in task-specific architectural integration and optimization, rather than in a fundamentally new algorithmic paradigm. Key components, such as dilated convolutions, SE-style feature recalibration, gated feature fusion, and lightweight design, are all closely related to existing ideas. Consequently, while the overall design is well targeted to the problem setting and appears empirically effective, the methodological novelty is more likely to be viewed as incremental rather than fundamental.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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 this paper addresses an important problem and presents positive experimental results, with a clear motivation and a structural design that achieves improvements in both detection performance and computational efficiency across multiple datasets, its overall contribution falls slightly below the acceptance threshold.
The main issue lies in the fact that the study’s contribution is primarily reflected in the structural optimization of the denoising U-Net backbone within an existing weakly supervised diffusion framework, rather than in advancing innovations in the weakly supervised mechanism or the diffusion formulation itself. Specifically, the method integrates direction-guided multi-scale modeling, feature recalibration based on the SE mechanism, gated skip-connection fusion, and lightweight convolution design—techniques that exhibit significant similarities to existing approaches. As a result, although this integrated solution is logical and well-aligned with the task, the methodological novelty may be considered insufficient.
Overall, the paper falls slightly below the acceptance criteria. However, if the authors can effectively address these concerns in their rebuttal, the final acceptance of the paper may still be considered.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The rebuttal is polite and well-intentioned, but fails to substantively address the most critical concern.
The key question that repeatedly raised by both the Meta-Reviewer and Reviewer #2 is whether the performance gains genuinely stem from the proposed DMS/ASSF/PDS-Conv architecture, or are merely artifacts of the AnoFPDM inference scheme. Their response “we will incorporate this into future work” is a deferral, not an answer.
On novelty, the authors argue their contribution lies in task-specific integration rather than in individual operations. This framing is defensible but does not change the fact that the core building blocks (directional dilated convolutions, SE attention, high-frequency gating via Gaussian subtraction, channel-shuffled lightweight convolutions) are all closely derived from prior work. No new theoretical insight or methodological breakthrough is presented.
My original concern about the lack of quantitative analysis by lesion type or boundary complexity was also left unaddressed.
In summary, the rebuttal clarifies some ablation details but sidesteps the most important question. I maintain my original recommendation: Weak Reject (3).
Review #2
- Please describe the contribution of the paper
The authors propose a modified diffusion-based architecture for weakly supervised medical anomaly detection, focusing on improving feature extraction within the denoising network. The method introduces two main components: (i) a Direction-guided Multi-Scale (DMS) module that enhances multi-scale and directional feature representation at the U-Net bottleneck, and (ii) an Anomaly-Sensitive Skip Fusion (ASSF) mechanism that gates skip connections using high-frequency structural cues.
Notably, ASSF computes a high-frequency residual via subtraction of a Gaussian-smoothed feature map, which is equivalent to a Laplacian-style decomposition. This biases the model toward edge and boundary information when propagating skip features.
These components are integrated into a lightweight U-Net-style backbone (AnoSSPD), and evaluated within the AnoFPDM forward diffusion anomaly detection framework.
- 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.
Through explicitly improving feature extraction, this directly impacts the quality of representations a diffusion anomaly detection model can reconstruct across low and high frequency regions. This is therefore a well motivated paper with strong empirical results and improves a frequently overlooked component in diffusion anomaly detection.
The paper is well motivated: improving feature extraction within diffusion models is an important and relatively under-explored direction for anomaly detection.
The method demonstrates strong empirical performance across multiple datasets.
The lightweight design is practically valuable, achieving significant reductions in computational cost while maintaining performance.
- 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.
The method is evaluated exclusively within the AnoFPDM forward-process framework, where anomaly detection is driven by differences between guided and unguided trajectories. As a result, it is unclear whether performance gains stem from improved generative modeling (i.e., a better learned score field) or from architectural biases that directly enhance anomaly sensitivity.
The evaluation would benefit from comparing against using this architecture for reconstruction approaches such as AnoDDPM. As the paper is written, it’s hard to assess the efficacy of the architecture away from the AnoPFDM inference scheme.
Note, it’s not really clear what the top layer of Table 2 is assessing. The numbers suggest it’s the AnoFPDM results from Table 1? A more complete ablation would lead to a better per-component assessment. While this reviewer doesn’t expect every permutation, the bottleneck DMS module is present in every ablation and could be swapped for the standard Conv block equivalent.
Further, how does this architecture compare with other popular UNet frameworks such as ConvNeXt (v1/v2)? - Obviously they would need to be tweaked to inject timestep embeddings at a minimum
The description of the AnoFPDM-based inference is brief. A short summary of the forward-process anomaly scoring (e.g., how sub-maps are generated and aggregated) would improve readability and make the paper more self-contained.
Figure 1 is quite dense and difficult to read with really small text. Splitting the figure up into separate figures such that text can be legibly read would improve clarity. Perhaps a figure for (a) and another for (b and c)? Tables 3 & 4 could be downsized to help account for this.
- 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
To indicate the difference in baselines, it would be clearer to place a horizontal line in table 1, indicating the unsupervised vs weakly supervised methods. This is listed in the text at the start of 4.2 but not overly clear at a glance.
Due to the presence of non-FLOP-dominant operations (e.g., squeeze-and-excitation), reporting inference time (e.g., latency or time per sample) would better reflect the practical efficiency of the model.
- 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 addresses an underexplored but important aspect of diffusion-based anomaly detection, the backbone architecture, and demonstrates strong empirical results alongside meaningful efficiency gains, making it a valuable contribution.
While the ablations provide some evidence that the architectural components are effective within the AnoFPDM framework, it remains unclear how well these improvements generalise beyond this setting (e.g., to reconstruction-based approaches or alternative backbones). Additional comparisons would help better contextualise the gains.
Overall, the paper fills a clear gap and shows promising results, but the evaluation limits how fully the contribution can be assessed. This reviewer leans toward acceptance.
- 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.
Ultimately, this reviewer maintains their decision from above that the paper addresses a highly overlooked component in unsupervised anomaly detection with diffusion models, the architecture.
This reviewer’s overall opinion is that the architecture has been compared against other methodologies but not other architectures. Although comparisons against alternative backbones (including pretrained checkpoints) would strengthen the paper, such experiments are not expected for rebuttal. However, the standard diffusion U-Net [Diffusion Models Beat GANs on Image Synthesis (2021)] remains the de facto backbone across many diffusion anomaly detection methods, improvements over AnoFPDM are still meaningful. To conclude, this reviewer feels the strengths (including the Tab. 2 ablation) of this paper marginally outweigh the limitations.
Lastly, this reviewer would implore the authors to tidy the clarity of the manuscript by briefly describing AnoFPDM and as Reviewer 3 raises (“What’s AnoDDPM (G|S)?”) to help the manuscript become more self-contained, and refer the reader to related work for more information.
Review #3
- Please describe the contribution of the paper
The proposed AnoSSPD addresses the challenges of complex lesion morphologies and high computational costs in weakly supervised medical anomaly detection. The new framework integrates the DMS module, ASSF mechanism, and a lightweight PDS-Conv backbone.
- 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.
The presented model endows the denoising network with explicit multi-scale field-of-view and directional structural awareness using DMS and ASSF. DMS decouples directional features and aggregates distinct receptive fields to supplement anatomical shape information across scales. ASSF utilizes a high-frequency structural prior gating mechanism to ensure the model efficiently focuses on fine details and edge morphology within anomalies.
- 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.
The title of the paper is “…Weakly Supervised Medical Image Anomaly Detection”, and there should be explanation why the method is “Weakly” supervised. The abbreviation of presented method is “AnoSSPD”, but there is no its full name in the paper. There are duplicated blocks in Fig1(a), and they should be simplified. What are c-ctx and h-dec used for? There is no explanation, and they are not used in formula. In the section of “3.4 Lightweight Network Architecture”, only the existing technique of Point-Depth Shuffle Convolution is introduced, thus the section is almost meaningless. In Table 1, the compared baselines include AnoDDPM(G/S), what is the difference between them? There are 6 methods for comparison in Table 1, why only 4 methods are used in Fig. 2? From Table 3, it can be found that the evaluation metrics become worse after using the lightweight architecture strategy.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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?
A new framework is presented, but the paper needs further modification and optimization.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
As the comments from different reviewers, there are still problems in the paper, and it needs further revision for publication.
Author Feedback
We thank all reviewers for your valuable comments for improving our work. To Meta Reviewer: In our Rebuttal, we have provide dclarifications on the methodological details and ablation studies to address the concerns raised. To Reviewer #1: Q1: There is a lack of analytical depth on specific lesion types, anatomical structures, or boundary complexities. A1: Regarding analytical depth, we clarify that the DMS module yields the most significant performance gains on irregular lesions with highly variable shapes and sizes, such as pediatric tumors in BraTS-PEDs and complex jagged stroke boundaries in ATLAS. Existing methods often keep generic U-Net denoisers lacking medical structural awareness, whereas we provide a morphology-aware denoising backbone. Q2: The methodological novelty is incremental. A2: Our novelty is not in claiming each operation is entirely new, but in redesigning and integrating them into a task-specific structural inductive bias based on the inherent deficiencies of current diffusion models for medical anomaly detection. Specifically, we emphasize three points: (1) Task-specific innovation: DMS is not a simple stacking of dilated convolutions, but a structural design oriented towards directional, multi-scale, and boundary-sensitive modeling of medical anomalous regions. (2) Framework adaptability innovation: The modules are embedded into the AnoFPDM forward-process anomaly scoring, aiming to enhance the capability of modeling differences between guided and unguided trajectories. (3) Efficiency-performance trade-off innovation: The lightweight structure is not merely compressing parameters, but reducing inference costs while maintaining anomaly detection performance. To Reviewer #2 Q1-Q2: The method is evaluated exclusively within AnoFPDM, making it hard to assess the efficacy of the architecture away from this inference scheme. A1-A2: We appreciate this insightful and highly necessary recommendation. We fully accept this valuable suggestion and will incorporate these cross-framework validations into our future work. Q3: The top layer of Table 2 is unclear. A3: All ablations in Table 2 keep the AnoFPDM forward inference, training protocol, and data split fixed, only changing backbone components. The first row is the original AnoFPDM result. The following rows show the effects of adding DMS to different layers and further incorporating ASSF. Thus, Table 2 is a controlled ablation showing that improvements are attributable to the proposed shape-scale-aware backbone components, not the AnoFPDM inference scheme alone. Q4-Q6: Other concerns including comparison with ConvNeXt, brief description of AnoFPDM, and dense Fig. 1.A4-6: We will incorporate cross-framework validations (like ConvNeXt) into our future work. Regarding presentation issues such as Fig. 1, please refer to our unified response to R3-Q2~Q7.To Reviewer #3 Q1: The evaluation metrics become worse in Table 3 after using the lightweight architecture. A1: We agree that our wording should be more precise. PDS-Conv is not intended to further improve accuracy, but to provide a practical trade-off between performance and computational cost for iterative diffusion inference. Compared with the full backbone, PDS-Conv causes only a slight performance drop, while reducing parameters from 97.89M to 46.12M and GFLOPs from 65.00 to 14.75.Thus, the lightweight version substantially reduces computational cost with only minor accuracy loss. In future versions, we will describe this more accurately. Q2-Q7: Presentation and clarity issues, including explanations for “Weakly” supervised, full name of AnoSSP, usage of c_ctx and h_dec, AnoDDPM(G/S), and figure readability. A2-7: We thank the reviewer for pointing these out. h_dec denotes the current decoder feature; c_ctx is obtained by global average pooling on h_dec to generate modulation parameters in ASSF. In future versions, we will fix these issues carefully and thoroughly.
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.
The empirical results are strong across three datasets and the efficiency gains are well-documented. The authors should address the most important points raised by the reviewers, especially with regard to whether the performance gains are attributable to the proposed architectural components or to the AnoFPDM inference scheme, the performance degradation under the lightweight architecture visible in Table 3, and the novelty positioning relative to existing directional convolution and gated skip connection designs.
- 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 paper addresses an underexplored aspect of diffusion-based anomaly detection, the denoising backbone architecture, with consistent empirical improvements across three datasets. The rebuttal adequately addressed the three main concerns raised: the within-framework ablation in Table 2 provides controlled evidence that gains are attributable to the proposed components rather than the AnoFPDM inference scheme alone, the lightweight architecture performance degradation is reframed appropriately as a compute-performance tradeoff with concrete efficiency numbers, and the novelty is positioned as task-specific structural integration. Cross-framework generalizability remains an acknowledged limitation that should be discussed in the final version.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
After considering the rebuttal and updated reviews, I recommend rejection. The rebuttal does not sufficiently resolve the concern that the observed gains may be difficult to attribute specifically to the proposed architectural components, and several components appear to be incremental combinations of existing operations. Given these unresolved concerns about contribution attribution and methodological novelty, I do not think the paper sufficiently crosses the acceptance threshold in its current form.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
While the proposed method demonstrates its superiority across three datasets, and the integration of existing components could be, up to some extent, seen as novel, there are important limitations that undermines the impact of the proposed work. First, I believe that the empirical validation is limited, with several important works on anomaly detection that appeared during 2025 missing (including papers at MICCAI’25). Furthermore, one of the critical issues is the concept of weakly supervised, which is not explained in the paper (rather referred to be explained in a further revision). Similarly, I found the use of the wording Shape a bit overstated: authors frame their approach as shape-aware, but there is nothing that supports why this is the case. Given all these concerns, in addition to several concerns from reviewers which remain unresolved, I recommend the rejection of this work in its current form.