List of Papers Browse by Subject Areas Author List
Abstract
Anatomical landmark detection is a fundamental task in medical image analysis supporting a wide range of diagnostic and interventional workflows. Although recent methods have achieved sub-millimetric localisation, accuracy alone is not sufficient for clinical deployment, requiring reliability and robustness in prediction. Despite its clinical relevance, the impact of representation learning in this context is still underexplored. In this work, we introduce CDPM-Align, a multi-scale guidance-aligned conditional diffusion pre-training for anatomical landmark detection. Our experimental setup focuses on a few images and a few annotation regimes. Specifically, we employ three popular heterogeneous small-scale benchmark datasets for representation learning via conditional generative pre-training. Furthermore, we consider low-annotation scenarios for the downstream task of landmark detection, with 10 and 25 annotated images, reflecting realistic trade-offs between clinical effort and resource constraints for annotations. Our results confirm that generative pre-training enables the model to learn a robust representation. This improves both accuracy and uncertainty on the downstream tasks, advancing towards safe and efficient clinical deployment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3797_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Malga-Vision/CDPM-Align
Link to the Dataset(s)
N/A
BibTex
@InProceedings{DiRob_CDPMAlign_MICCAI2026,
author = { Di Via, Roberto AND Voiculescu, Irina AND Odone, Francesca AND Pastore, Vito Paolo},
title = { { CDPM-Align: Multi-Scale Guidance-Aligned Diffusion Pretraining for Robust Few-Shot Anatomical Landmark Detection } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The primary contribution of CDPM-Align is a novel conditional diffusion pretraining framework designed to achieve robust anatomical landmark detection in low-annotation clinical regimes. The core methodological innovation is a multi-scale alignment objective that leverages the classifier-free guidance signal as an explicit structural descriptor, enforcing feature consistency across different diffusion timesteps and UNet hierarchy levels. This mechanism enables the model to learn stable, dataset-specific representations from small-scale heterogeneous datasets, significantly improving both localization accuracy and spatial reliability while maintaining robustness against the structural variability.
- 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 paper introduces a novel technical mechanism that uses classifier-free guidance signals as structural descriptors, ensuring robustness to anatomical heterogeneity where standard SSL methods often fail. The framework demonstrates high data efficiency, matching the performance of of large model using a targeted in-domain corpus of 988 images. This approach is backed by rigorous multi-metric validation across three benchmarks. Ultimately, the results confirm that the model maintains clinical precision and stability in 10-shot and 25-shot scenarios.
- 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 proposed method is currently restricted to 2D X-ray images, leaving its performance and the scalability of the multi-scale alignment loss unproven for 3D volumetric data such as CT or MRI. Furthermore, the framework relies on heuristic hyperparameter selection; the alignment weights and diffusion timesteps are determined empirically without a formal theoretical derivation. From a resource perspective, the alignment phase introduces a 1.3x computational overhead due to the requirement of four forward passes per iteration, which may pose a barrier for hardware-constrained environments. Finally, the evaluation lacks a direct head-to-head comparison with emerging medical foundation models, instead relying on externally reported figures rather than standardized few-shot benchmarks.
- 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 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.
(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 introduces a technically sound and novel method for exploiting diffusion signals as structural descriptors, effectively solving the representation collapse often seen in discriminative SSL for medical imaging. The framework’s high data efficiency is particularly impressive, achieving state-of-the-art results in extreme 10-shot and 25-shot regimes using minimal in-domain data. While currently limited to 2D modalities and reliant on some heuristic tuning, this contribution still well-suited for the MICCAI community.
- Reviewer confidence
Somewhat confident (2)
- [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 authors propose CDPM-align, a conditional diffusion pre-training framework for anatomical landmark detection, with a novel multi-scale guidance alignment loss. The method is evaluated in low-data settings (10 shot and 25 shot fine-tuning) across three datasets. Results indicate improved performance in both mean radial error (MRE) and predictive uncertainty (ERE) compared to baseline approaches, suggesting the effectiveness of the proposed strategy.
- 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.Novel multi-scale guidance alignment loss The authors introduce a multi-scale guidance alignment loss that enforces directional consistency of the classifier-free guidance signal (Delta h) across independently sampled diffusion steps. This design is well-motivated and demonstrates measurable improvements over baseline methods in the experimental results.
2.Clinically motivated evaluation setting The work focuses on challenging low-data settings (10 shot and 25 shot landmark detection), which are highly relevant to real-world clinical scenarios where annotations are scarce. This strengthens the practical significance of the study.
3.Computation-aware design The proposed strategy activates the multi-scale alignment loss only during a later training phase (10% of total iterations). This is a thoughtful design choice that significantly reduces computational overhead (from 4× to 1.3×) while maintaining strong performance gains. (Sec. 2.2)
- 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.Details for baseline initialization and training protocol In Section 3.1, the setup for the baseline models is not sufficiently clear, leading to potential architectural confusion. Specifically, it is unclear if the authors:
(a) Constructed a U-Net using a pre-trained ResNet-101 encoder (e.g. DINO) and then performed the Stage 1 generative pre-training (diffusion) for the baseline, or; (b) Simply utilized the pre-trained ResNet-101 encoder (e.g., DINO) to pre-train the pooled corpus and attached a randomly initialized decoder only at the beginning of the Stage 2 downstream task.
The manuscript would be significantly strengthened if this initialization protocol was explicitly clarified.
2.Baseline fairness If the (b) from the point 1 is true, the proposed CDPM-align model possesses a significant structural advantage by starting the 10-shot task with a fully pre-trained decoder, whereas the baselines would be forced to learn dense upsampling from scratch. This is supported by the fact that all self-supervised ResNet-101 baselines (DINO, MoCo v3, SimCLRv2) in Tables 1-3 consistently underperform compared to ResNet-101 (ImageNet). On the same task (while different dataset), prior work Di Via et al [8] has shown self-supervised ResNet-101 encoders (DINO, MoCo v3, SimCLRv2) outperform ImageNet-supervised ResNet-101.Yet in this paper, the opposite pattern appears. A possible explanation is that the baselines are disadvantaged by the decoder initialization protocol. I suggest clarifying the baseline model construction and re-assess the fairness of the evaluation.
3.Uncertainty quantification claims Although the proposed method adopts the Expected Radial Error (ERE) as a measure of uncertainty, the uncertainty analysis remains insufficient. In particular, given the authors’ claim that a lower ERE indicates better uncertainty quantification, a more thorough analysis and discussion of this relationship would be expected.
4.(Minor) Readability of figures and tables While the experimental evaluation is comprehensive, all figures and tables are difficult to read due to small font sizes, particularly in printed format. The authors are encouraged to improve readability and ensure compliance with standard conference formatting guidelines (i.e., 8 pt minimum font size, but the font size in the Table 1-3 seems to be smaller than 8 pt ). I encouraged authors to double check to avoid possible desk rejection.
- 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 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.
(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 experimental setup, particularly the baseline initialization and training protocol, is insufficiently described. This raises concerns regarding the fairness and interpretability of the reported comparisons. Additionally, some presentation issues (e.g., the readability of figures and tables) further limit the overall clarity of the manuscript. Overall, while the scientific contributions and empirical results are promising, I recommend a weak accept, provided that the authors clarify the baseline setup and improve the readability of the manuscript.
- 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 #3
- Please describe the contribution of the paper
This paper proposes a conditional diffusion pretraining framework with multi-scale guidance alignment (CDPM-align) for few-shot anatomical landmark detection, with a focus on improving both accuracy and uncertainty estimation.
- 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.Strong and comprehensive experiments. The method is evaluated across multiple datasets, metrics (MRE, ERE, SDR, P95), and low-shot settings, showing consistent improvements.
2.Clinically relevant focus on uncertainty. Incorporating ERE and demonstrating improved uncertainty concentration is a meaningful and less explored aspect.
3.Good performance in low-data regimes. The method shows clear advantages under 10/25-shot settings and strong data efficiency.
4.Consistent gains over diverse baselines. Improvements are observed across supervised, SSL, and diffusion-based methods.
- 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.Limited conceptual novelty. The approach mainly combines existing ideas (diffusion pretraining, classifier-free guidance, consistency alignment) with moderate innovation.
2.Insufficient justification of design choices. The alignment mechanism (multi-scale, cosine similarity, timestep selection) is only heuristically motivated and lacks deeper analysis.
- 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 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.
(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?
Despite moderate novelty, the paper is well-motivated, methodologically sound, and supported by strong experimental evidence. The consistent improvements in low-shot settings and the focus on uncertainty make it a valuable contribution, especially for clinically relevant applications.
- 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 all reviewers for the encouraging feedback. We are glad the guidance alignment mechanism is recognised as novel and well-motivated (R2, R3, R4), and that the few-shot evaluation (R3, R4) and experimental evidence (R2, R4) are appreciated. We address the remaining concerns below, following the submitted paper’s reference numbering. Camera ready includes the link to our GitHub repository, where we release the code as open-source.
Baseline initialisation and fairness (R3-1, R3-2). The SSL baselines use a UNet with a pretrained ResNet-101 encoder and a randomly initialised decoder; the ImageNet baseline uses the same architecture. CDPM/CDPM-align pretrains the full UNet via diffusion, replacing only the output head downstream. SSL vs. ImageNet is architecturally matched (both share random decoders), so SSL underperformance points to encoder representation quality rather than decoder disadvantage. The discrepancy with [8], where SSL outperformed ImageNet, stems from two differences: (i) here, SSL is pretrained on the pooled multidomain corpus (chest + cephalo + hand) rather than a single-domain setting (as done in [8]), with discriminative objectives collapsing on heterogeneous data; (ii) this work adopts contour-hugging heatmaps [19] instead of Gaussian ones as in [8], a more representation-sensitive setting. Unlike SSL, the generative paradigm pretrains the entire UNet by construction; the ablation in Table 4 (lambda=0 vs. lambda=5) isolates the alignment contribution within the same architecture.
Novelty and design choices (R4-1, R4-2, R2). We appreciate R4’s recognition that the paper is ”well-motivated, methodologically sound, and supported by strong experimental evidence”. The novelty lies in repurposing the classifier-free guidance signal ∆h as a structural descriptor for representation learning. To our knowledge, no prior work has enforced cross-timestep directional consistency on ∆h for dense prediction. Table 4 confirms that alignment alone improves MRE by 24% on Shenzhen 10-shot. Cosine similarity is magnitude-invariant, suited for ∆h whose norm varies across timesteps while its direction encodes structural information. The range [T/4, 3T/4] targets the regime where ∆h is most class-discriminative, avoiding both near-clean and noise-dominated timesteps. Multiscale extraction follows established UNet feature hierarchy practice [1].
Uncertainty quantification (R3-3). Uncertainty quantification is clinically relevant (highlighted as a strenght by R4). Under this perspective, we exploited uncertainty estimation as an evaluation metric for different pre-training strategies. Accordingly, we relied on ERE as a quantitative measure of uncertaintly, as validated originally in [19], which introduced the metric and showed its correlation with MRE under contour-hugging heatmaps. ERE measures the expected distance between the predicted landmark and a heatmap sample, capturing spatial concentration. Crucially, ERE is informative only when heatmaps allow freeform distributions, as in [19]; Gaussian targets with fixed shape make ERE a trivial function of bandwidth. CDPM-align achieves ERE at most equal to MRE across all 6 conditions, with sub-millimetre ERE at 25-shot on ISBI2015 and DHA (25-34% improvement over [8]).
2D limitation and overhead (R2). We agree that 3D extension is important future work (Sec. 4); the alignment mechanism is dimensionality-agnostic. The 1.3x overhead, which R3 acknowledges as a ”thoughtful design choice”, adds 20 min on A100.Regarding medical foundation models, we incorporated a concurrent work [11] in our paper and benchmarked our few-shot results against the fully supervised results reported in its tables, obtaining comparable performance and supporting the robustness of our guidance-alignment signal in generative pretraining.
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.
I agree with the reviewers on the contribution and the novelty of this paper.