Revisiting Integration of Image and Metadata for DICOM Series Classification: Cross-Attention and Dictionary Learning

Matthias Lenga; Melanie Dohmen; Sara Lorio; Tuan Truong

arxiv: 2602.23833 · v2 · pith:6BE4XIBSnew · submitted 2026-02-27 · 📡 eess.IV · cs.CV

Revisiting Integration of Image and Metadata for DICOM Series Classification: Cross-Attention and Dictionary Learning

Tuan Truong , Melanie Dohmen , Sara Lorio , Matthias Lenga This is my paper

Pith reviewed 2026-05-22 10:50 UTC · model grok-4.3

classification 📡 eess.IV cs.CV

keywords DICOM series classificationmultimodal learningcross-attentiondictionary learningmedical imagingmissing metadataimage fusion

0 comments

The pith

A multimodal framework fuses DICOM images and metadata using cross-attention and dictionary learning to classify series without imputation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces an end-to-end multimodal method for identifying DICOM image series that handles missing metadata and variable lengths. It encodes images with a 2.5D approach and metadata with a sparse dictionary encoder, then fuses them via bi-directional cross-modal attention. This setup is tested on liver MRI datasets where it outperforms image-only, metadata-only, and other multimodal baselines. The approach aims to support large-scale medical image analysis by improving robustness to real-world data issues like incomplete acquisition metadata.

Core claim

The proposed method jointly models image content and acquisition metadata for DICOM series classification by using bi-directional cross-modal attention for fusion and a sparse, missingness-aware dictionary encoder for metadata, without requiring imputation, leading to consistent outperformance over baselines in in-domain and out-of-domain evaluations.

What carries the argument

Bi-directional cross-modal attention mechanism fused with a sparse missingness-aware dictionary encoder based on learnable feature dictionaries and value-conditioned modulation.

If this is right

Improved robustness for DICOM series classification in presence of missing or inconsistent metadata.
Consistent superiority over image-only and metadata-only approaches across evaluation settings.
Handling of variable series lengths through equidistant slice sampling in a 2.5D visual encoder.
Direct applicability to quality control and protocol harmonization in large medical image datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This could lead to fewer errors in downstream tasks like automated analysis pipelines.
Similar dictionary learning approaches might apply to other domains with sparse multimodal data.
Future work could test the method on additional imaging modalities beyond MRI.

Load-bearing premise

That integrating bi-directional cross-modal attention with the sparse missingness-aware dictionary encoder will improve classification robustness without any imputation or bias-inducing preprocessing.

What would settle it

Demonstrating that the proposed method does not outperform relevant baselines on a new dataset with high rates of missing metadata would falsify the claim of improved robustness.

Figures

Figures reproduced from arXiv: 2602.23833 by Matthias Lenga, Melanie Dohmen, Sara Lorio, Tuan Truong.

**Figure 1.** Figure 1: Proposed method: pixel data of S DICOM slices is embedded in visual feature pathway. DICOM metadata is embedded by the Sparse Metadata Encoder. Bidirectional cross-modal attention contextualizes all image and metadata embeddings. Final integration to a series-level representation is done by learnable pooling. phase, it is inherently unreliable. Metadata fields are vendor-dependent, frequently manually e… view at source ↗

**Figure 2.** Figure 2: In-domain evaluation: five-fold cross-validation per-class F1 scores (%) on the Duke Liver MRI dataset. Concatenation baselines (4) and (5) use S = 3 slices as inputs. (4) uses zero metadata imputation while (5) uses a small MLP to predict missing fields based on observed values and learnable baseline feature values. We apply a consistent preprocessing scheme to the raw DICOM header values. Categorical tag… view at source ↗

read the original 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 on the publicly available Duke Liver MRI dataset and a large multi-institutional in-house cohort, assessing 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a targeted multimodal fix for DICOM series classification that skips imputation by using a learnable dictionary encoder plus bi-directional cross-attention, but the abstract supplies no numbers or ablations so the outperformance claim stays unverified.

read the letter

This paper gives a workable way to classify DICOM series by fusing images and metadata while skipping imputation for missing fields, but the lack of detailed results makes it hard to tell how much better it really is. The core idea pairs a 2.5D image branch with a sparse dictionary-based metadata encoder that modulates on available values, then links them through bi-directional cross-attention. They test on the Duke Liver MRI set and a multi-institutional in-house cohort, claiming gains over image-only, metadata-only, and other multimodal baselines in both in-domain and out-of-domain settings. Variable series length is handled by equidistant slice sampling rather than padding or cropping tricks. What stands out is the explicit design for missing metadata without any fill-in step, which avoids one common source of bias in real hospital data. The architecture description is clear and the choice to evaluate generalization across institutions is a reasonable step. The soft spots sit in the evidence. The abstract asserts consistent outperformance yet shows no metrics, error bars, dataset sizes, or ablation tables. Without those, it is difficult to isolate whether the dictionary encoder or the cross-attention actually drives the reported edge, especially when metadata is sparse or absent. The stress-test note is on point here: if the model simply falls back to the image branch in low-metadata cases, the multimodal advantage could be narrower than claimed. This work is aimed at people building automated pipelines for medical image quality control or protocol harmonization who deal with messy DICOM headers every day. A reader who needs practical robustness in multimodal fusion for healthcare data would find the architecture details useful. It deserves a serious referee because the problem is common and the method is a focused extension of existing techniques rather than a vague idea. I would send it for review but ask for quantitative results, metadata-completeness breakdowns, and ablations that remove the dictionary or attention components one at a time.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes an end-to-end multimodal framework for DICOM series classification. It encodes images via a 2.5D visual encoder on equidistantly sampled slices and metadata via a sparse missingness-aware encoder based on learnable feature dictionaries with value-conditioned modulation (no imputation). These are fused with bi-directional cross-modal attention. The approach is evaluated on the Duke Liver MRI dataset and a large multi-institutional in-house cohort, with claims of consistent outperformance over image-only, metadata-only, and multimodal 2D/3D baselines in both in-domain and out-of-domain settings.

Significance. If the empirical results are shown to be driven by the proposed components, the work could meaningfully advance robust classification of heterogeneous DICOM series in medical imaging pipelines, particularly by avoiding imputation biases when metadata is missing or inconsistent. The explicit handling of series-length variability and cross-modal interactions addresses real practical challenges, though the overall significance hinges on demonstrating that the dictionary encoder and cross-attention provide gains beyond standard fusion.

major comments (2)

[§4] §4 (Experimental Results), Table 2 and Table 3: The central claim of consistent outperformance across all settings and robustness to missing metadata requires evidence that the sparse missingness-aware dictionary encoder and bi-directional cross-attention are the drivers of gains. No ablation is reported that isolates or removes the dictionary component (or the cross-attention) while keeping other modeling choices fixed, nor is performance stratified by metadata completeness. This omission is load-bearing because the abstract emphasizes handling entirely missing metadata by design; without these controls it remains possible that gains are dataset-specific or arise from the 2.5D image branch alone.
[§3.2] §3.2 (Metadata Encoder): The value-conditioned modulation and learnable dictionaries are presented as jointly enabling robustness without imputation, yet the manuscript provides no targeted experiment or analysis for the case of completely absent metadata. If the encoder reduces to a learned default vector in that regime, the multimodal advantage over image-only baselines could be illusory, directly undermining the robustness claim.

minor comments (2)

The abstract would be strengthened by including at least one key quantitative result (e.g., accuracy or AUC delta with dataset size) to support the outperformance claim.
[§3.2] Notation for the dictionary learning and modulation (e.g., how missing values are masked before modulation) could be formalized with a short equation or pseudocode for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments. We address each major comment below and indicate the revisions planned for the next manuscript version.

read point-by-point responses

Referee: [§4] §4 (Experimental Results), Table 2 and Table 3: The central claim of consistent outperformance across all settings and robustness to missing metadata requires evidence that the sparse missingness-aware dictionary encoder and bi-directional cross-attention are the drivers of gains. No ablation is reported that isolates or removes the dictionary component (or the cross-attention) while keeping other modeling choices fixed, nor is performance stratified by metadata completeness. This omission is load-bearing because the abstract emphasizes handling entirely missing metadata by design; without these controls it remains possible that gains are dataset-specific or arise from the 2.5D image branch alone.

Authors: We agree that isolating the contributions of the dictionary encoder and cross-attention, along with stratification by metadata completeness, would strengthen the claims. In the revised manuscript we will add ablations that (i) replace the dictionary encoder with a standard metadata encoder while keeping the 2.5D image branch and cross-attention fixed, and (ii) replace bi-directional cross-attention with simple concatenation while keeping the dictionary encoder fixed. We will also report performance broken down by metadata completeness (fully present, partially missing, and entirely missing) on both the Duke and in-house cohorts. revision: yes
Referee: [§3.2] §3.2 (Metadata Encoder): The value-conditioned modulation and learnable dictionaries are presented as jointly enabling robustness without imputation, yet the manuscript provides no targeted experiment or analysis for the case of completely absent metadata. If the encoder reduces to a learned default vector in that regime, the multimodal advantage over image-only baselines could be illusory, directly undermining the robustness claim.

Authors: The sparse missingness-aware encoder is explicitly constructed so that an all-missing metadata input maps to a learned default vector that is still passed through value-conditioned modulation and then fused via cross-attention; this is not equivalent to simply dropping the metadata branch. To directly test the concern, the revised manuscript will include an experiment that forces complete metadata absence on both evaluation cohorts and reports the resulting multimodal versus image-only performance gap. revision: yes

Circularity Check

0 steps flagged

No derivation chain or self-referential reductions present

full rationale

The paper describes an empirical multimodal neural architecture (bi-directional cross-attention plus sparse missingness-aware dictionary encoder) evaluated on external datasets (Duke Liver MRI and multi-institutional cohort). No equations, first-principles derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. Performance claims rest on direct comparisons to image-only, metadata-only, and other multimodal baselines, which are externally falsifiable and not forced by construction or internal definitions.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on domain assumptions about the effectiveness of the proposed neural modules rather than new mathematical derivations or external benchmarks.

axioms (2)

domain assumption Bi-directional cross-modal attention can effectively fuse image and metadata features for classification.
Invoked in the description of the fusion mechanism.
domain assumption Sparse learnable feature dictionaries can process metadata with missing or inconsistent values without imputation.
Central to the metadata encoder design.

pith-pipeline@v0.9.0 · 5775 in / 1194 out tokens · 44496 ms · 2026-05-22T10:50:08.075342+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages · 1 internal anchor

[1]

Journal of Digital Imaging36(1), 289–305 (Feb 2023)

Cluceru, J., Lupo, J.M., Interian, Y., Bove, R., Crane, J.C.: Improving the Automatic Classification of Brain MRI Acquisition Contrast with Ma- chine Learning. Journal of Digital Imaging36(1), 289–305 (Feb 2023). https://doi.org/10.1007/s10278-022-00690-z, https://doi.org/10.1007/s10278-022- 00690-z

work page doi:10.1007/s10278-022-00690-z 2023
[2]

Journal of Digital Imaging33(3), 747–762 (Jun 2020)

Gauriau, R., Bridge, C., Chen, L., Kitamura, F., Tenenholtz, N.A., Kirsch, J.E., Andriole, K.P., Michalski, M.H., Bizzo, B.C.: Using DICOM Metadata for Ra- diological Image Series Categorization: a Feasibility Study on Large Clinical Brain MRI Datasets. Journal of Digital Imaging33(3), 747–762 (Jun 2020). https://doi.org/10.1007/s10278-019-00308-x, https:...

work page doi:10.1007/s10278-019-00308-x 2020
[3]

Densely connected convolutional networks,

Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely Connected Convolutional Networks. In: 2017 IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR). pp. 2261–2269 (Jul 2017). https://doi.org/10.1109/CVPR.2017.243, https://ieeexplore.ieee.org/document/8099726, iSSN: 1063-6919

work page doi:10.1109/cvpr.2017.243 2017
[4]

Academic Radiology32(3), 1192–1203 (Mar 2025)

Kim, B., Mathai, T.S., Helm, K., Mukherjee, P., Liu, J., Sum- mers, R.M.: Automated Classification of Body MRI Sequences Us- ing Convolutional Neural Networks. Academic Radiology32(3), 1192–1203 (Mar 2025). https://doi.org/10.1016/j.acra.2024.11.046, https://linkinghub.elsevier.com/retrieve/pii/S1076633224008912

work page doi:10.1016/j.acra.2024.11.046 2025
[5]

Scientific Reports 15(1), 27044 (Jul 2025)

Kim, J., Chae, A., Duda, J., Borthakur, A., Rader, D.J., Gee, J.C., Kahn, C.E., Witschey, W.R., Sagreiya, H.: Automated characteriza- tion of abdominal MRI exams using deep learning. Scientific Reports 15(1), 27044 (Jul 2025). https://doi.org/10.1038/s41598-025-11985-w, https://www.nature.com/articles/s41598-025-11985-w

work page doi:10.1038/s41598-025-11985-w 2025
[6]

Frontiers in Neuroinformatics15(Nov 2021)

Liang, S., Beaton, D., Arnott, S.R., Gee, T., Zamyadi, M., Bartha, R., Symons, S., MacQueen, G.M., Hassel, S., Lerch, J.P., Anagnostou, E., Lam, R.W., Frey, B.N., Milev, R., Müller, D.J., Kennedy, S.H., Scott, C.J.M., Investigators, T.O., Strother, S.C., Troyer, A., Lang, A.E., Greenberg, B., Hudson, C., Corbett, D., Grimes, D.A., Munoz, D.G., Munoz, D.P....

work page doi:10.3389/fninf.2021.622951 2021
[7]

Radiology: Artificial Intelligence5(5), e220275 (Sep 2023)

Macdonald, J.A., Zhu, Z., Konkel, B., Mazurowski, M.A., Wiggins, W.F., Bashir, M.R.: Duke Liver Dataset: A Publicly Available Liver MRI Dataset with Liver Segmentation Masks and Series Labels. Radiology: Artificial Intelligence5(5), e220275 (Sep 2023). https://doi.org/10.1148/ryai.220275, https://pubs.rsna.org/doi/full/10.1148/ryai.220275

work page doi:10.1148/ryai.220275 2023
[8]

Abdominal Radiology49(10), 3735–3746 (Oct 2024)

Miller, C.M., Zhu, Z., Mazurowski, M.A., Bashir, M.R., Wiggins, W.F.: Automated selection of abdominal MRI series using a DICOM metadata classifier and selective use of a pixel-based classifier. Abdominal Radiology49(10), 3735–3746 (Oct 2024). https://doi.org/10.1007/s00261-024-04379-5, https://doi.org/10.1007/s00261-024- 04379-5 10 Truong et al

work page doi:10.1007/s00261-024-04379-5 2024
[9]

FiLM: Visual Reasoning with a General Conditioning Layer

Perez, E., Strub, F., Vries, H.d., Dumoulin, V., Courville, A.: FiLM: Visual Reasoning with a General Conditioning Layer (Dec 2017). https://doi.org/10.48550/arXiv.1709.07871, http://arxiv.org/abs/1709.07871, arXiv:1709.07871 [cs]

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1709.07871 2017
[10]

Truong, T., Mohammadi, S., Lenga, M.: How Transferable are Self- supervised Features in Medical Image Classification Tasks? In: Proceed- ings of Machine Learning for Health. pp. 54–74. PMLR (Nov 2021), https://proceedings.mlr.press/v158/truong21a.html

work page 2021
[11]

In: Advances in Neural Information Processing Systems

Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, L., Polosukhin, I.: Attention is All you Need. In: Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)

work page 2017
[12]

Current Medical Imag- ing Formerly Current Medical Imaging Reviews21, e15734056361649 (Sep 2025)

Yuan, C., Jia, X., Wang, L., Yang, C.: Fine-grained Prototype Network for MRI Sequence Classification. Current Medical Imag- ing Formerly Current Medical Imaging Reviews21, e15734056361649 (Sep 2025). https://doi.org/10.2174/0115734056361649250717162910, https://www.eurekaselect.com/244037/article

work page doi:10.2174/0115734056361649250717162910 2025
[13]

In: Advances in Neural Infor- mation Processing Systems

Zaheer, M., Kottur, S., Ravanbakhsh, S., Poczos, B., Salakhutdi- nov, R.R., Smola, A.J.: Deep Sets. In: Advances in Neural Infor- mation Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://papers.nips.cc/paper_files/paper/2017/hash/f22e4747da1aa27e363d86d40ff442fe- Abstract.html

work page 2017
[14]

IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 1688–1698 (Apr 2022)

Zhu, Z., Mittendorf, A., Shropshire, E., Allen, B., Miller, C., Bashir, M.R., Mazurowski, M.A.: 3D Pyramid Pooling Network for Abdominal MRI Series Classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 1688–1698 (Apr 2022). https://doi.org/10.1109/TPAMI.2020.3033990, https://ieeexplore.ieee.org/document/9242262

work page doi:10.1109/tpami.2020.3033990 2022

[1] [1]

Journal of Digital Imaging36(1), 289–305 (Feb 2023)

Cluceru, J., Lupo, J.M., Interian, Y., Bove, R., Crane, J.C.: Improving the Automatic Classification of Brain MRI Acquisition Contrast with Ma- chine Learning. Journal of Digital Imaging36(1), 289–305 (Feb 2023). https://doi.org/10.1007/s10278-022-00690-z, https://doi.org/10.1007/s10278-022- 00690-z

work page doi:10.1007/s10278-022-00690-z 2023

[2] [2]

Journal of Digital Imaging33(3), 747–762 (Jun 2020)

Gauriau, R., Bridge, C., Chen, L., Kitamura, F., Tenenholtz, N.A., Kirsch, J.E., Andriole, K.P., Michalski, M.H., Bizzo, B.C.: Using DICOM Metadata for Ra- diological Image Series Categorization: a Feasibility Study on Large Clinical Brain MRI Datasets. Journal of Digital Imaging33(3), 747–762 (Jun 2020). https://doi.org/10.1007/s10278-019-00308-x, https:...

work page doi:10.1007/s10278-019-00308-x 2020

[3] [3]

Densely connected convolutional networks,

Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely Connected Convolutional Networks. In: 2017 IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR). pp. 2261–2269 (Jul 2017). https://doi.org/10.1109/CVPR.2017.243, https://ieeexplore.ieee.org/document/8099726, iSSN: 1063-6919

work page doi:10.1109/cvpr.2017.243 2017

[4] [4]

Academic Radiology32(3), 1192–1203 (Mar 2025)

Kim, B., Mathai, T.S., Helm, K., Mukherjee, P., Liu, J., Sum- mers, R.M.: Automated Classification of Body MRI Sequences Us- ing Convolutional Neural Networks. Academic Radiology32(3), 1192–1203 (Mar 2025). https://doi.org/10.1016/j.acra.2024.11.046, https://linkinghub.elsevier.com/retrieve/pii/S1076633224008912

work page doi:10.1016/j.acra.2024.11.046 2025

[5] [5]

Scientific Reports 15(1), 27044 (Jul 2025)

Kim, J., Chae, A., Duda, J., Borthakur, A., Rader, D.J., Gee, J.C., Kahn, C.E., Witschey, W.R., Sagreiya, H.: Automated characteriza- tion of abdominal MRI exams using deep learning. Scientific Reports 15(1), 27044 (Jul 2025). https://doi.org/10.1038/s41598-025-11985-w, https://www.nature.com/articles/s41598-025-11985-w

work page doi:10.1038/s41598-025-11985-w 2025

[6] [6]

Frontiers in Neuroinformatics15(Nov 2021)

Liang, S., Beaton, D., Arnott, S.R., Gee, T., Zamyadi, M., Bartha, R., Symons, S., MacQueen, G.M., Hassel, S., Lerch, J.P., Anagnostou, E., Lam, R.W., Frey, B.N., Milev, R., Müller, D.J., Kennedy, S.H., Scott, C.J.M., Investigators, T.O., Strother, S.C., Troyer, A., Lang, A.E., Greenberg, B., Hudson, C., Corbett, D., Grimes, D.A., Munoz, D.G., Munoz, D.P....

work page doi:10.3389/fninf.2021.622951 2021

[7] [7]

Radiology: Artificial Intelligence5(5), e220275 (Sep 2023)

Macdonald, J.A., Zhu, Z., Konkel, B., Mazurowski, M.A., Wiggins, W.F., Bashir, M.R.: Duke Liver Dataset: A Publicly Available Liver MRI Dataset with Liver Segmentation Masks and Series Labels. Radiology: Artificial Intelligence5(5), e220275 (Sep 2023). https://doi.org/10.1148/ryai.220275, https://pubs.rsna.org/doi/full/10.1148/ryai.220275

work page doi:10.1148/ryai.220275 2023

[8] [8]

Abdominal Radiology49(10), 3735–3746 (Oct 2024)

Miller, C.M., Zhu, Z., Mazurowski, M.A., Bashir, M.R., Wiggins, W.F.: Automated selection of abdominal MRI series using a DICOM metadata classifier and selective use of a pixel-based classifier. Abdominal Radiology49(10), 3735–3746 (Oct 2024). https://doi.org/10.1007/s00261-024-04379-5, https://doi.org/10.1007/s00261-024- 04379-5 10 Truong et al

work page doi:10.1007/s00261-024-04379-5 2024

[9] [9]

FiLM: Visual Reasoning with a General Conditioning Layer

Perez, E., Strub, F., Vries, H.d., Dumoulin, V., Courville, A.: FiLM: Visual Reasoning with a General Conditioning Layer (Dec 2017). https://doi.org/10.48550/arXiv.1709.07871, http://arxiv.org/abs/1709.07871, arXiv:1709.07871 [cs]

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1709.07871 2017

[10] [10]

Truong, T., Mohammadi, S., Lenga, M.: How Transferable are Self- supervised Features in Medical Image Classification Tasks? In: Proceed- ings of Machine Learning for Health. pp. 54–74. PMLR (Nov 2021), https://proceedings.mlr.press/v158/truong21a.html

work page 2021

[11] [11]

In: Advances in Neural Information Processing Systems

Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, L., Polosukhin, I.: Attention is All you Need. In: Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)

work page 2017

[12] [12]

Current Medical Imag- ing Formerly Current Medical Imaging Reviews21, e15734056361649 (Sep 2025)

Yuan, C., Jia, X., Wang, L., Yang, C.: Fine-grained Prototype Network for MRI Sequence Classification. Current Medical Imag- ing Formerly Current Medical Imaging Reviews21, e15734056361649 (Sep 2025). https://doi.org/10.2174/0115734056361649250717162910, https://www.eurekaselect.com/244037/article

work page doi:10.2174/0115734056361649250717162910 2025

[13] [13]

In: Advances in Neural Infor- mation Processing Systems

Zaheer, M., Kottur, S., Ravanbakhsh, S., Poczos, B., Salakhutdi- nov, R.R., Smola, A.J.: Deep Sets. In: Advances in Neural Infor- mation Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://papers.nips.cc/paper_files/paper/2017/hash/f22e4747da1aa27e363d86d40ff442fe- Abstract.html

work page 2017

[14] [14]

IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 1688–1698 (Apr 2022)

Zhu, Z., Mittendorf, A., Shropshire, E., Allen, B., Miller, C., Bashir, M.R., Mazurowski, M.A.: 3D Pyramid Pooling Network for Abdominal MRI Series Classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 1688–1698 (Apr 2022). https://doi.org/10.1109/TPAMI.2020.3033990, https://ieeexplore.ieee.org/document/9242262

work page doi:10.1109/tpami.2020.3033990 2022