List of Papers Browse by Subject Areas Author List
Abstract
Zero-shot anomaly localisation via vision-language models (VLMs) offers a compelling approach for rare pathology detection, yet current methods suffer from a lack of healthy reference context, high prediction variance, and suboptimal reference selection. We introduce WALDO, a training-free framework grounded in optimal transport theory that addresses these limitations through: (i) entropy-weighted Sliced Wasserstein distance for anatomically-aware reference selection from DINOv3 patch distributions, (ii) Goldilocks zone sampling exploiting the non-monotonic relationship between reference similarity and localisation accuracy, and (iii) self-consistency aggregation via weighted non-maximum suppression. On the NOVA brain MRI benchmark, WALDO with Qwen2.5-VL-72B achieves 43.5 ± 1.6% mAP@30 (95% CI: [40.4, 46.7]), representing a 19% relative improvement over zero-shot baselines. Cross-model evaluation shows consistent gains: GPT-4o achieves 32.0 ± 6.5% and Qwen3-VL-32B achieves 32.0 ± 6.6% mAP@30. Paired McNemar tests confirm statistical significance for the primary pair (p = 1.8 × 10⁻⁶, hit@30). We provide theoretical analysis of the Goldilocks effect through distributional divergence, demonstrating why moderate-similarity references optimise the bias-variance tradeoff in comparative visual reasoning. Code: https://github.com/bkainz/WALDO_MICCAI26_demo
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0474_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/bkainz/WALDO_MICCAI26_demo
Link to the Dataset(s)
NOVA brain MRI benchmark (rare-pathology localisation): https://huggingface.co/datasets/c-i-ber/Nova
VinDr-CXR chest X-ray dataset (radiologist-annotated lesion bounding boxes): https://physionet.org/content/vindr-cxr/1.0.0/
BibTex
@InProceedings{KaiBer_WassersteinAligned_MICCAI2026,
author = { Kainz, Bernhard AND Müller, Johanna P. AND Baugh, Matthew M. G. AND Bercea, Cosmin},
title = { { Wasserstein-Aligned Localisation for VLM-Based Distributional OOD Detection in Medical Imaging } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a zero-shot anomaly localization with vision-language models (VLMs), using an entropy-weighted Sliced Wasserstein distance over DINOv2 patch distributions, samples “Goldilocks” references of moderate similarity via a DPP scheme, and aggregates VLM predictions using a self-consistency weighted NMS. The approach show competitive and remarked results.
- 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 use of Sliced Wasserstein distances over patch embeddings is well-motivated for capturing distributional and spatial structure -NOVA experiments are thorough, with multiple VLMs and ablations showing consistent gains. -Framing zero-shot anomaly localization as comparative inference with healthy references is clinically intuitive -The method is promising for triage and attention guidance
- The comparative prompting design leverages VLMs’ strengths for differential reasoning, an underexplored angle in medical anomaly localization.
- 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 Goldilocks hypothesis is compelling and supported by stratification, but the “theory” is a qualitative bias–variance narrative -There is necessary a Statistical robustness study. Statistical test that evidence significant differences. The paper should specify the binary criterion and justify the test choice
- No comparison against reference-based CLIP variants
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
The paper is well motivated, clear and the methodology is novel. The problem is challenging and face a current problem. However, there is demanding a major clarity about statistical analysis, selection criteria and test choice. Also a major explanation about bias-variance and the Goldilocks hypothesis . If possible, it is demanding a baseline comparison with CLiP variants.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes that moderately similar (30th–70th percentile) healthy references optimize the VLM’s bias-variance tradeoff for anomaly detection. The framework introduces entropy-weighted sliced wasserstein distance for anatomically-aware patch-level alignment between patient and reference scans. The authors formulate zero-shot localisation as a comparative inference problem using healthy in-context anatomical priors.
- 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.Training-free framework that enables general-purpose VLMs to perform specialized medical tasks without needing labeled anomaly datasets or fine-tuning. 2.The paper leverages sliced wasserstein distance (O(nlogn)) to preserve spatial structure better than global cosine similarity. 3.The method demonstrates consistent performance gains across multiple VLM architectures including GPT-4o, Qwen2.5, and Gemini-2.0-Flash.
- 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.Reported mAP@30 remains lower than vanilla Qwen3-VL-235B, questioning the utility of the added pipeline complexity. 2.The “Goldilocks” concept mirrors existing “identity mapping” prevention strategies in atlas-based segmentation and generative models. 3.Using healthy references for comparison is the foundational principle of radiology and frameworks like PatchCore or WinCLIP. 4.The figure mistakenly references a non-existent DinoV3 and lacks clear connectivity between selected references and query. 5.Does not compare the Goldilocks strategy against standard sampling techniques like K-Means or Coreset selection. 6.Fails to address how volumetric MRI data is utilized, likely losing critical Z-axis spatial information. 7.Utilizing five references per query increases VLM token count and latency by 5x, hindering real-time deployment. 8.The computational cost of patch-level Sliced Wasserstein across thousands of healthy images is not clearly detailed. 9.Final cumulative gains are not clearly aggregated in a single table against Cosine or SW2 strategies. (clearer tab 3)
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
Please address the topics raised in weakness section
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The authors propose WALDO - a training-free framework utilizing optimal transport theory to localise pathologies in brain MRI and chest X-ray images. The framework processes a query image and a healthy reference pool with DINOv2, extracting patch embeddings and measuring distributional similarity through an entropy-weighted Sliced Wasserstein distance. References from the “Goldilocks zone” are selected via DPP for diversity. Performance is evaluated on two datasets; for the NOVA brain MRI benchmark, the authors also reproduced the benchmark’s baseline results. WALDO significantly outperforms the baselines, achieving improvements of more than +19% mAP@30 on both brain MRI and chest X-ray images.
- 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 proposed framework is training-free, allowing VLMs to perform in a zero-shot regime. 2.The authors provide a theoretical analysis of the non-monotonic relationship between reference similarity (needed to correctly identify a healthy reference image based on anatomical consistency) and localization accuracy (needed to correctly localise the pathology). 3.The paper is well-written and the code in the repository is well-organised, reflecting a strong commitment to reproducibility. Notably, the authors reproduced the results from the benchmark and provided a thorough ablation study evaluating the contribution of every component of the method (Table 3). 4.The sensitivity analysis of hyperparameters is a valuable addition, demonstrating that the authors are aware of the importance of robustness to parameter choices.
- 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.Page 2, “Limitations of Current Approaches.” In L3, the authors state that stochastic sampling causes high prediction variance with standard deviations exceeding 5% mAP — it is unclear how this value was obtained. A reference or derivation should be provided. 2.Page 4, “Implementation Details.” The authors mention that images were resized to 512×512, but the initial image resolution is never stated. Additionally, λ = 0.1 is used, whereas intuitively a value of 0.5 might seem more natural. A more detailed justification of how these values were chosen would strengthen the paper. 3.It is unclear why only 200 samples were selected from the VinDr-CXR dataset. A justification for this choice should be provided, as it may affect the reliability of the evaluation. 4.In Table 1, Qwen3-VL-235B (MoE) outperforms WALDO. While the authors attribute this to the model’s expert routing mechanism conflating reference and query contexts, it remains unclear whether the performance gap is simply attributable to the significantly larger model size. This leaves an important comparison underexplained. 5.In Table 1, five models are considered; however, Qwen3-VL-235B (MoE) and Gemini-2.0-Flash are absent from further experiments, such as Table 2.The rationale for this inconsistent model selection should be clarified. 6.Page 5, “Experiments. Models.” The authors mention that ablation experiments were also conducted with DINOv3, but the corresponding results are nowhere to be found in the paper or in the repository. These should either be included or the reference should be removed. 7.Page 6, “Experiments. Models.” The paper states that Table 2 shows chest X-ray performance in the range of 24–35% mAP@30, whereas the actual values in Table 2 range from 10.9 to 34.1.This inconsistency should be corrected. Furthermore, the claim that heart size and shape vary substantially among healthy individuals is used to explain lower chest X-ray performance, but it is not discussed whether similar variability exists in brain MRI, which would undermine this explanation. 8.Page 7, “Experiments. Sensitivity Analysis.” The sensitivity analysis covers only hyperparameter values in the vicinity of the chosen ones. It remains unclear whether other, more distant values could yield better performance, which limits the conclusions that can be drawn from this analysis. 9.The authors mention VLM hallucinations, but the frequency and impact of them in the experiments is never discussed. Given that the framework relies on VLMs, this is a notable omission that should be addressed. 10.The paper lacks any comparison with supervised methods on the same tasks, or at least an explanation of why such a comparison would be impossible or inappropriate. Including such a comparison, even indicatively, would help contextualise the achieved performance and provide a clearer picture of the remaining gap between training-free and fully supervised approaches.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
1.The VinDr-CXR dataset is not mentioned in the abstract, despite being used in the evaluation. It would be worth at least briefly noting its inclusion. 2.The title of Table 3 should specify that Qwen2.5-VL-72B was used, to avoid ambiguity. 3.Page 6, “Experiments. Goldilocks Effect Analysis.” It should be explicitly noted that the results discussed in this section correspond to Table 1, to help the reader follow the analysis more easily.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents WALDO, a training-free framework for unsupervised pathology localisation in medical images that operates in a zero-shot regime, removing the need for annotated training data - a significant bottleneck in medical imaging. The experimental results demonstrate strong performance, with improvements exceeding +19% mAP@30 over the baselines on both brain MRI and chest X-ray images.
While several weaknesses were identified - including small inconsistencies in reported metrics and underexplained design choices - these are largely addressable through revisions and do not undermine the core contributions of the work. The major weaknesses raised should be addressed before finalisation, but none of them represent fundamental flaws. Overall, the strengths of the paper outweigh its weaknesses, and I recommend acceptance subject to the authors addressing the comments provided.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers and AC for the constructive feedback. The camera-ready paper addresses every point compatible with the “no new experiments / no number changes” rule. We summarise the updates and clarify a few misinterpretations.
Statistical robustness (R1.W2, R3.W9, AC). The criterion is now stated explicitly: paired McNemar tests on the binary hit@30 (IoU>=0.3) indicator at alpha=0.05.The Qwen2.5-VL-72B NOVA pair p-value is tightened from “p<0.01” to the exact p=1.8e-6 (same paired test, same sample size). §Error Analysis adds a hallucination quantification (R3.W9): 27-47% of WALDO predictions on VinDr-CXR are false positives (IoU=0), yet WALDO reduces the per-image no-correct-localisation rate by 22-55 percentage points compared to zero-shot, so the recall gain dominates the false-positive volume.
Backbone naming (R3.W6, R2.W4). The submitted text inconsistently referred to “DINOv2”; all experiments were in fact run with DINOv3-ViT-B/16 (facebook/dinov3-vitb16-pretrain-lvd1689m). Abstract, Figure 1 caption, and §Models are corrected. No reported numbers change. Post-hoc analysis revealed that there is no measureable difference between v2 and v3 for our method.
Factual corrections (R3.W3, R3.W7, INT). §Implementation now states n=949 for VinDr-CXR (the “n=200” sentence was a stale leftover; Table 2 was already n=949). The CXR mAP@30 prose range “24-35%” is corrected to “10.9-34.1%” to match Table 2.NOVA sample count is corrected from 907 to 906.Image resolutions (NOVA 480x480, VinDr-CXR up to 1024x1024 DICOM) and the lambda=0.1 grid-search rationale are now stated (R3.W2). The +360% relative-gain figure for GPT-4o on CXR was a derived-number typo and is corrected to +230% (recomputed from Table 2; (10.9-3.3)/3.3).
Qwen3-VL-235B regression (R2.W1, R3.W4). §Results now explicitly rules out model scale: Qwen2.5-VL-72B is ~3x smaller yet has comparable zero-shot performance (36.4%) and benefits strongly from WALDO (+19.5%). The regression appears MoE-specific.
Clarifications. (i) (R2.W2) The Goldilocks effect is distinct from identity-mapping prevention: we operate on patch-level feature distributions via Sliced Wasserstein at prompt time, exploiting a bias-variance tradeoff in differential VLM reasoning. The mechanism and granularity differ from latent-replay / regularisation strategies in atlas-based or generative methods. (ii) (R3.W7) §Results now states explicitly that brain-MRI variability exists, but focal pathologies produce contrast changes absent in healthy tissue, making healthy references more discriminative for MRI than for CXR. (iii) (R2.W6, AC) NOVA provides pre-extracted 2D axial slices; the reported experiments do not require volumetric handling. (iv) (R3.W10) §Limitations explains the absence of supervised baselines: WALDO targets the setting where labelled training data is absent, which is the norm for rare diseases.
Items deferred to the post-publication extended version. Under the no-new-experiments rule, the following could not be added to the camera-ready and are flagged for an extended arXiv and journal version: a WinCLIP / CLIP-based reference baseline (R1.W3, R2.W3, AC); K-Means / Coreset reference-selection ablations (R2.W5); CXR runs of Qwen3-VL-235B and Gemini-2.0-Flash (R3.W5); a wider hyperparameter sweep beyond the chosen-vicinity (R3.W8); and a more formal information-theoretic treatment of the Goldilocks bound (R1.W1). We thank the reviewers for these strong directions.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 paper introduces a zero-shot method for finding anomalies in medical images with vision-language models and without the need for labeled training data . The idea of using Wasserstein distance to compare image patches and picking healthy references that are moderately similar to the query is clever and deserves a pass.
Despite my provisional accept recommendation, I agree that the main hypothesis appears to lack rigorous statistical testing and feels more like a qualitative story than a proven theory (which is fine, but the paper needs to be presented accordingly). Also, the paper needs to compare against a simple CLIP-based baseline or a standard sampling strategy. Lastly, the authors should address the latency increase and make some clarification to experimental details (3D MRI data handing, etc., Also: why only 200 samples from the chest X-ray dataset?