List of Papers Browse by Subject Areas Author List
Abstract
Automated identification of DICOM image series is essential for large-scale medical image analysis, quality control, protocol harmonization, and reliable downstream processing. However, DICOM series classification remains challenging due to heterogeneous slice content, variable series length, and entirely missing, incomplete or inconsistent DICOM metadata. We propose an end-to-end multimodal framework for DICOM series classification that jointly models image content and acquisition metadata while explicitly accounting for all these challenges. (i) Images and metadata are encoded with modality-aware modules and fused using a bi-directional cross-modal attention mechanism. (ii) Metadata is processed by a sparse, missingness-aware encoder based on learnable feature dictionaries and value-conditioned modulation. By design, the approach does not require any form of imputation. (iii) Variability in series length and image data dimensions is handled via a 2.5D visual encoder and attention operating on equidistantly sampled slices. We evaluate the proposed approach for classifying liver MRI DICOM series, using the publicly available Duke Liver MRI dataset and a large, multi-institutional in-house cohort, and assess both in-domain performance and out-of-domain generalization. Across all evaluation settings, the proposed method consistently outperforms relevant image-only, metadata-only and multimodal 2D/3D baselines. The results demonstrate that explicitly modeling metadata sparsity and cross-modal interactions improves robustness for DICOM series classification.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3515_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/MatthiasLen/img-meta-cls
Link to the Dataset(s)
Duke dataset: https://pmc.ncbi.nlm.nih.gov/articles/PMC10546360/
BibTex
@InProceedings{TruTua_Revisiting_MICCAI2026,
author = { Truong, Tuan AND Dohmen, Melanie AND Lorio, Sara AND Lenga, Matthias},
title = { { Revisiting Integration of Image and Metadata for DICOM Series Classification: Cross-Attention and Dictionary Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors present a multiodal DICOM series classification framework. In their manuscript, the authors want to automate the identification of MRI DICOM series, which is often problematic due to missing metadata information. The uahtors state, that while there are both methods focused on the metadata, as well as on the image data, a combined approach has not yet been proposed.
A 2.5D image encoder is combined with a missingness-aware metadata encoder as well as bi-directional cross-attention. The presented approach is tested on the Duke Liver MRI dataset and a larger in-house dataset.
- 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 problem addressed in the paper is relevant in real-world clinical settings and the authors motivate their work properly. 2) The proposed framework is also well motivated and understandable. 3) The reported results are promising. 4) It is nice to see abation studies. Even though there could have been more ablation studies regarding design choices, the 8-page limit is probably the main reason for the lack of further ablations. 5) The out-of-domain experiment strengthens the approach further.
- 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 authors seem to frame their work as a general DICOM series classification framework, however, all experiments are limited to liver MRI. This should be made more clear, perhaps even in the title. This limitation is mentioned in the discussion, but again, it should be mentioned already at the beginning of the manuscript.
2) More in-depth analyses or explanation of why the performance drops for some classes on the out-of-domain dataset would be of interest (e. g. Dixon opposed-phase or portal venous)
3) The proposed model is more complex than the mentioned baselines. A comparison in inference time, model parameter cound or FLOPs is desriable.
4) The authors present a multimodal framework. When it comes to multimodality, the interplay and influence of the different modalities is of interest. Are there labels in the metadata that are more important than others? When does the image information become more important? Were these things tested / analyzed?
5) Why were the more recent works, which were also cited in the paper, not implemented as baselines (E. g. Kim et al. 2025 or Yuan et al. 2025)? If these models were not applicable, a short statement in the manuscript would be good.
- 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
The paper is clearly written and well-organized. My only point would be the term “Revisiting” in the title. What exactly is being revisted in this work?
- 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 motivation is sound and the results are promising. I am looking forward to the response of the authors to my mentioned questions in “weaknesses”. I will revise my scoring upwards depending on the answers in the rebuttal phase.
- 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 describes a novel multi-modal framework combining DICOM metadata and image slides information to classify the series using contextualization (other neighbor series) DICOM metadata missingness is handled by using a sparse metadata encoder and avoids using missing value imputation (DICOM header data is often incomplete). It is evaluated on a large multi-centric in-house MRI dataset and a publicly available dataset to perform in-domain and out-of-domain evaluation.
- 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.
Paper clearly written and easy to follow. Interesting practical topic that offers a very accurate and interesting robust approach for MRI sequences classification. This is often a very important step in PACS data retrieval and first step in data cleaning or MRI sequences routing for downstream AI tools or research. The fact that the approach is robust to data sparsity is very interesting. The paper offers a thorough comparison to existing approaches and with/wo metadata with/wo image information.
- 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.
# Contextualisation / introduction Quantitate evidence from previous research that metadata alone is not enough? % contribution MRI-sequences classification => not clearly mentioned (classification could also be anatomy-based for instance) Not clear if this framework is totally new or was taken from another domain. Multimodal frameworks are used across many different applications.
# Literature review not complete. Some examples of missing references: Deep Learning–driven classification of external DICOM studies for PACS archiving (JOnske et al., 2022) https://d-nb.info/1266272461/34
Automatic sequence identification in multicentric prostate multiparametric MRI datasets for clinical machine-learning (de Almeida et al., 2025) https://pmc.ncbi.nlm.nih.gov/articles/PMC12187622/
# Dataset
- “We merge closely related early arterial, mid arte- rial, and late arterial into a unified Arterial class, and merge late and transi- tional phases into a single Late class, yielding 13 classes overall. “ => this choice should be justified (not enough data or performance not good enough?). We can think of downstream applications that would require this information. is the dataset balanced?
- what anatomy is included?
- what about MRI DWI B-values?
- how was the ground-truth determined?
- what’s in the “Other” class?
- what about FLAIR sequences?
# Modeling
- How were the DICOM attributes selected? Why 116? Did they only select Type I and II DICTOM attributes? What about private tags that depend on the manufacturer?
- The rapidity of training is not discussed although it would be interesting to know how long it takes as we can imaging such approach would have to be retrained on new centers to avoid domain-shift.
# Evaluation / interpretation:
- DICOM metadata are prone to change over time (change in protocol, new machines etc.), temporal generalisation is an important component of such approaches and it is not evaluated here.
- In and Out-of-domain evaluation limited to liver MRI only => could it be evaluated on other anatomy? (there are plenty of publicly available MRI datasets out there)
- DWI b-values are important in liver imaging, could they be included in the final output?
- Out-of-domain shift lacks of analysis insights. For the drop on Dixon opposed-phase for instance, it would have been great to explain the failures and give more insights on why it was failing.
- Tables 2 and 3 are missing confidence intervals to truly compare results
- The contribution of the sparse missingness-aware metadata encoder is not analyzed thoroughly. Could be interesting to mention the % of missing data in both datasets to get a sense of how much data is imputed.
- 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?
Important topic that often overlooked in the medical imaging community. This is a very fundamental step for cleaning up the medical imaging data for research / models training but also when integrating AI models in production. The solution proposed is interesting and seems to be very accurate.
- 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 paper presents a model combining metadata and image data for DICOM series classification, an important problem given the notable unreliability of the DICOM series description tag.
- 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.
- Paper addresses an important practical problem
- Methodological contribution is solid
- 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.
- Choice of evaluation metrics is questionable
- Limited scope for evaluation
- No source code or pre-trained model is available
- 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.
(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 model architecture is sound and the methodological contribution is solid. The main comments I have on the paper relate to the evaluation.
1.The evaluation includes a large set of DICOM series but it is limited to liver MRI. Given the scope covered by the DICOM specification, even just in MRI, it would have been very useful to report an evaluation on a broader set of images, e.g., brain, heart, and possibly modalities (e.g., CT, PET).
2.The evaluation metrics are limited to global (weighted) and per-class F1 scores (and precision and recall). Given the likely data imbalance (not reported but very likely in such a dataset), it would be important to report other metrics including balanced accuracy, and ROC-AUC. I’m also uncertain if the differences reported in Figure 2 are statistically significant.
In addition, no link is available for source code or pre-trained model. Reproducibility would be very difficult to assess.
- 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
Author Feedback
We thank all reviewers for their constructive feedback and for recognizing the clinical relevance of our work. We address the major concerns below: (1) Novelty vs. metadata-only and multimodal works (Meta, R1, R2) While our work is based on self-attention, we propose a DICOM-specific end-to-end multimodal design to address: (i) sparse metadata with missingness and (ii) variable-length 3D series that require cross-slice context. Prior DICOM series classification works in our discussion are either metadata-only, image-only, or a two-stage / late fusion pipeline that are not missingness-aware and do not learn to model joint representation through interaction of cross-modality and cross-slice context. Our pipeline deals with missing data without relying on any kind of imputation mechanism, due to the embedding dictionary design in the encoder path.
(2) Clarification for baseline choices (Meta, R1, R2) We isolate the contributions of sparse metadata encoder (SME) and bidirectional cross-attention (BCA) by comparing against unimodal, two-stage, or standard fusion approaches. To keep this comparison interpretable, we include one representative standard 2D/3D image-only CNN, trained with the same preprocessing as the multimodal approach. Kim et al. (2025) and Yuan et al. (2025) present variants of image-only networks, making it harder for us to attribute gains to vision network choices or our proposed SME/BCA components. As a result, we did not include these approaches as direct comparators. We will clarify the rationale behind the baseline choices as well as expand related work in the final version.
(3) Evaluation scope is limited to liver MRI (Meta, R1-3) We acknowledge that our experimental results are restricted to only liver MRI. Validating the performance on additional anatomies and modalities would require curated series labels with consistent mappings across datasets to enable in-domain and out-of-domain evaluation. This scope will be stated explicitly in the final version.
(4) Metrics and uncertainty reporting (Meta, R1-3) We reported the mean and std over 5-fold CV and tested the significance using the Wilcoxon signed-rank test (p < 0.05) (Sec 4.1). We chose the weighted F1 score as primary target, due to class imbalance and multi-class setup, in line with prior work and enabling future benchmarking.
(5) Failure analysis for certain classes in out-of-domain evaluation (Meta, R1-2) Out-of-domain drops are concentrated in a few classes (Table 3), notably Dixon opposed-phase and portal venous. The confusion matrix shows: (i) many Dixon opposed-phase errors are assigned to OTHER class (i.e., non-targeted sequence types in the in-house schema), where the appearance may overlap with fat-suppressed T1 variants in the Duke protocol (ii) portal venous are often confused with the adjacent arterial phase, suggesting cross-site difference in contrast timing and protocol definitions. As discussed in Sec 4.2, we hypothesize that label-schema mapping can influence performance of specific subclasses.
(6) Interaction between modalities (Meta, R1) In Table 2 and Fig. 2, we provide evidence on how each modality contributes via unimodal and multimodal baselines. In particular, image-only models substantially outperform the metadata-only baseline, and multimodal fusion improves over image-only and over simple concatenation with fixed/learned imputation. This aligns with our discussion/limitations: metadata contribution is stronger for targets with well-populated tags (e.g., acquisition plane/MRCP), whereas missing or inconsistent fields shift the burden to the image pathway (Sec. 5).
(7) Reproducibility (Meta, R1-3) We will release code upon acceptance to support reproducibility. The link to the associated git repository will be added to the manuscript.
We hope these clarifications address the main concerns and accurately reflect the contributions and evidence already present in the submission.
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.
This paper presents a novel multimodal framework for identifying DICOM MRI series, aiming to address unreliable metadata. Overall, the reviewers agreed that the paper was well written, the motivation was clearly stated, and the topic was of high interest to the MICCAI community and beyond, in clinical settings. However, several concerns and questions have been raised. I believe that some of them might be addressed before the final version, and clearly encourage the authors to do so:
- R1 and R2 raised concerns about the novelty of the proposed framework compared to existing approaches. Recent works have been cited in the paper, but it is not clear how this particular framework improves beyond metadata-only methods and other multimodal frameworks.
- All reviewers found the evaluation limited in datasets and metrics. While an ablation study has been performed, some points remain unclear: – How were the baseline competitors chosen? – Why does the model fail on some classes more than others? – Which component contributes the most to performance? – Why did you only use liver MRI?
- R1 and R2 suggested adding more metrics on runtime and number of parameters.
- R2 and R3 pointed out the lack of statistical evidence on the results.
- R3 raised concerns about the reproducibility, and would kindly ask you to share the code and pretrained models. I recommend early acceptance but clearly recommend the authors to address these concerns in the final version.