MR-CLIP: Efficient Metadata-Guided Learning of MRI Contrast Representations

Jorge Cardoso; Mehmet Yigit Avci; Mehmet Yigitsoy; Paul Wright; Pedro Borges; Sebastien Ourselin

REVIEW 3 major objections 4 minor 20 references

MR-CLIP: Efficient Metadata-Guided Learning of MRI Contrast Representations

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MR-CLIP learns MRI contrast representations by aligning images with their DICOM metadata in a shared embedding space, without manual labels.

desk verdict Solid method paper with a genuine subject-leakage problem in the internal split; the external OASIS evaluation keeps the contribution afloat, but the generalization claim needs a subject-disjoint redo. read the letter →

arxiv 2507.00043 v2 pith:5HFW5SXO submitted 2025-06-23 cs.CV cs.AI

Mehmet Yigit Avci , Pedro Borges , Paul Wright , Mehmet Yigitsoy , Sebastien Ourselin , Jorge Cardoso This is my paper

classification cs.CVcs.AI

keywords contrastivelearningMRIcontrastrepresentationDICOMmetadatasupervisedlosscross-modalretrievalself-supervisedmedicalimaging

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that MRI contrast can be learned directly from the acquisition metadata that already accompanies every clinical scan, without any manual labelling. Because contrast in MRI is determined by settings like echo time, repetition time, and scanner sequence, the paper treats each distinct combination of these DICOM values as a class and trains a multimodal contrastive model to pull together slices with the same class and push apart slices with different classes. If this works, the resulting embeddings separate contrast while ignoring anatomy, which would make it possible to identify, retrieve, and organise brain scans in large heterogeneous hospital archives where sequence labels are missing or unreliable.

What carries the argument

The load-bearing object is the contrast-aware label: a unique combination of DICOM metadata values, with categorical tags matched exactly and numerical tags discretised so that similar acquisition settings collapse into one class. The loss that does the work is a supervised contrastive loss (SupCon) applied symmetrically in both directions, from image to text and text to image; it treats every slice sharing a label as a positive, allowing many-to-many supervision rather than strict one-to-one image-text pairing. The text prompts are generated from a standardised sentence template, which lets a language encoder consume the structured metadata. The grouping granularity acts as the main knob: coarser grids make retrieval easier and transfer across discretisations, while finer grids preserve more contrast detail.

What would settle it

Collect expert-labelled contrast labels (T1, T2, FLAIR, PD, etc.) for a sample of scans and check concordance with MR-CLIP's automatically assigned labels: any two scans that look identical but receive different MR-CLIP labels, or look different but share a label, would undermine the claim that the metadata grouping defines contrast.

Watch

Extended reading notes

Core claim

The central claim is that the continuous space of MRI contrast can be organised by the structured acquisition parameters stored in DICOM headers. MR-CLIP converts a selected set of categorical tags (used verbatim) and numerical tags (echo time and repetition time jointly quantised onto a grid, inversion time binned separately) into natural-language prompts, then trains a vision encoder and a text encoder with a symmetric supervised contrastive loss. All slices whose metadata groups match are positives for one another, so the model tolerates small acquisition differences and becomes invariant to subject and anatomy; different groups are pushed apart. The paper demonstrates that this yields effective cross-modal retrieval and accurate linear classification, and that the model transfers to an unseen dataset.

Load-bearing premise

The load-bearing premise is that the selected DICOM fields and their discretisation capture all visually meaningful contrast variation, so that slices sharing a label truly look alike and slices with different labels truly differ.

Editorial extensions

If this is right

Contrast-aware representations can be produced from any large hospital archive with DICOM headers, removing the need for expert contrast labels or protocol standardisation.
Cross-modal retrieval becomes possible: a 2D slice or a whole 3D scan can be matched to its acquisition metadata, and a metadata query can retrieve matching images.
Frozen MR-CLIP embeddings support linear classification of acquisition parameters, and the embeddings transfer from a fine grid to coarser contrast groupings without retraining.
Because training pulls together slices of the same contrast across different subjects and anatomies, the learned space should support downstream tasks that need anatomy-invariant, modality-aware features, such as data harmonisation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

One extension not pursued in the paper is to use the model's own predictions to detect metadata errors: slices whose predicted metadata systematically disagree with their DICOM header are candidates for corrupted or mislabeled tags.
The same recipe should transfer to other modalities whose appearance is set by acquisition parameters (CT, PET), provided a comparable structured label space can be defined from their headers.
The grouping granularity can be read as a regulariser; a data-adaptive discretisation learned from the embedding geometry might outperform fixed grids, especially for rare contrasts.
An implicit consequence is that the ceiling of the approach is set by metadata quality, so hospital-side cleaning and standardisation of DICOM headers would directly raise the quality of the learned representations.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Leif, here's the short version: this is a well-executed method paper with a real evaluation weakness that likely inflates the main numbers. The scan-level split with ~4.2 volumes per subject means the same subject's scans almost certainly appear in both train and test, so the 78.7% 3D scan-to-text retrieval and 82.6% linear accuracy partly reflect the model recognizing the subject rather than generalizing contrast knowledge. That's the most important thing to know.

What's genuinely new: the combination of CLIP-style multimodal alignment with SupCon loss over discretized DICOM metadata groups, plus within-volume positives, is not something I've seen. The scale is substantial — 40k subjects, 169k volumes, a range of scanners and protocols. The writing is clear, the design choices are explained, and the authors are upfront about the metadata-noise assumption. The external OASIS evaluation is a good idea: it is subject-disjoint, so it's the least leak-prone evidence, and it shows non-trivial retrieval even across different acquisition characteristics. That prevents me from dismissing the whole approach.

The weak spots are the ones you'd expect. First, no error bars or multiple runs, so we can't tell if differences between models are real. Second, the linear probe and part of the retrieval use the same grouped metadata labels that define the SupCon positives, so those numbers partly just confirm the training objective was optimized. Third, the internal split issue is serious; a subject-disjoint split would be the obvious fix. The paper claims anatomy-invariant representations, but the current evaluation doesn't rule out the alternative that the model exploits subject-specific features. The OASIS numbers, while lower, do offer some support for the generalization claim.

Who is this for? Anyone working on contrast-aware medical image representations or metadata-guided learning. It's a constructive counterpoint to label-hungry methods, and the released code and weights make it a useful baseline. I'd send it to review with a request for a subject-level split and error bars. If those come back clean, this could be a solid contribution; if not, the main claims need to be tempered.

Referee Report

3 major / 4 minor

Summary. This paper proposes MR-CLIP, a contrastive learning method for MRI that aligns 2D slices with structured DICOM metadata converted into text prompts. The method replaces the standard one-to-one InfoNCE objective with a supervised contrastive loss (Eq. 1) in which positive pairs share a discretized 'contrast-aware label' constructed from scanner, sequence, TE/TR, TI, and other DICOM tags. Training is performed on 169,634 clinical volumes from 40,005 subjects. Experiments report cross-modal retrieval (image-to-text, 3D scan-to-text, text-to-image) and linear probe classification, comparing against BiomedCLIP and an InfoNCE baseline, along with an out-of-distribution evaluation on OASIS (Table 3). The main claimed contributions are a scalable metadata-guided representation, scan- and within-volume contrast supervision, and a grid-based discretization strategy for label construction.

Significance. If the results hold, MR-CLIP would be a practically useful method for learning contrast-aware MRI representations without manual labels, and the public release of code and weights is a valuable contribution. The use of a large, diverse clinical dataset and the inclusion of an external OASIS evaluation are strengths, as is the interpretable grid-based grouping of acquisition parameters. The central limitation is that the internal evaluation is compromised by a likely subject-level leakage, and the OASIS results, while genuinely out-of-distribution, only partially compensate for that. The method itself is plausible and the contribution is meaningful, but the generalization claim needs stronger empirical support before publication.

major comments (3)

[Section 2 (Dataset)] The train/validation/test split is performed at the scan level on a dataset of 169,634 volumes from 40,005 subjects. With an average of about 4.2 volumes per subject, a random scan-level split will place volumes of the same subject in both the training and test sets: for a subject with four volumes and a 30% test fraction, the probability that at least one volume lands in the test set while at least one other does not is roughly 0.75. The image encoder can therefore exploit subject-specific anatomy or scanner-specific appearance to predict metadata at test time, rather than learning transferable, contrast-specific information. This threatens the internal validity of the retrieval and linear-probe results in Tables 1 and 2 and the claims of 'generalizable' and 'anatomy-invariant' representations. The authors should re-run the experiments with a subject-disjoint split, report subject and volume counts for each split, and state whether the main conclusions in Tables 1 and 2 are preserved.
[Eq. (1) and Section 3 (Evaluation)] The evaluation protocol is partly circular with respect to the training signal. The linear probe and the retrieval tasks are scored against the same grouped metadata labels that define the positive sets P(i) in the SupCon loss of Eq. (1). Consequently, high internal accuracy in Table 1 largely demonstrates that the model learned the training objective, not that the representation transfers to genuinely new contrast semantics. The OASIS experiments in Table 3 provide a valuable external check, but they still construct labels from the same kind of DICOM tags and do not test against independent expert annotations or held-out protocol families. Please add an evaluation on acquisition configurations not seen during training (for example, a protocol-disjoint split) or on expert-labeled sequence types, and state explicitly which tags were available and how the discretization grid was mapped when building the OASIS label space.
[Tables 1 and 2] All numbers in Tables 1 and 2 come from a single training run, with no standard deviations, confidence intervals, or significance tests. In Table 2 the differences between neighboring grid sizes are often only a few points (for example, 20x10 vs. 10x10 for image-to-text R@5: 80.4 vs. 84.3), so it is unclear whether the stated ranking of grouping strategies is stable. Please report means and standard deviations over multiple seeds for the main comparisons, or provide a clear justification for a single run given the computational budget.

minor comments (4)

[Section 3, first paragraph] There is a typo: 'significantly outperforming the the fine-tuned BiomedCLIP' should read 'significantly outperforming the fine-tuned BiomedCLIP'.
[Section 3, Grouping Strategy of Numerical Tags] The sentence 'This suggests that reducing label granularity simplify the alignment task' has a subject-verb agreement error; 'simplify' should be 'simplifies'.
[Implementation Details] The text encoder is not described; please specify the pretrained language model used to embed the metadata prompts. Table 1 also introduces a ViT-S/16 variant that is not explained in the method section.
[Section 4 (Discussion)] The statement that MR-CLIP 'decouples contrast information from anatomical content' is stronger than what the experiments directly measure; consider softening it or adding a quantitative anatomy-invariance or disentanglement evaluation.

Circularity Check

2 steps flagged · score 5.0 of 10

In-house retrieval and linear-probe numbers largely verify the training objective; OASIS provides independent support, so circularity is partial.

fitted input called prediction [Section 2 (Eq. 1) and Section 3 (Table 1, linear classification)]
""Each contrast-aware label corresponds to a unique combination of these metadata values." ... "where P(i) is the set of all other embeddings in the batch that share the same grouped contrast label as anchor i" ... "linear probe classification, where a simple linear classifier is trained on frozen image embeddings to predict metadata labels.""

The 'metadata labels' predicted by the linear probe are exactly the grouped contrast labels used to define P(i) in Eq. (1). Because the SupCon loss explicitly contracts embeddings of samples sharing a label and repels those with different labels, the reported 82.6% linear accuracy largely certifies that the training objective was optimized rather than demonstrating independent discovery of contrast structure. This is a fitted input being reported as a prediction.
fitted input called prediction [Section 3, Cross-modal Retrieval and Table 1]
""MR-CLIP employs supervised contrastive (SupCon) loss [7] that supports many-to-many positives... treats all samples sharing the same grouped contrast label as positives" and retrieval is evaluated as "retrieving metadata from 2D slices (image-to-text)... and retrieving images from metadata (text-to-image).""

The retrieval targets and queries are the same metadata-derived texts used as the supervised alignment signal in Eqs. (1)-(2). The final loss explicitly optimizes image-to-text and text-to-image SupCon, so high in-house R@1 on these tasks is a direct measure of how well the training objective was fit on the test split, not an independent test of generalization. The OASIS experiments are the non-circular evidence.

full rationale

MR-CLIP's central in-house evidence partly reduces to its own training objective. The SupCon loss in Eq. (1) defines positive pairs as slices sharing the same grouped metadata label, and the final loss Eq. (2) explicitly optimizes both image-to-text and text-to-image alignment. The linear probe and cross-modal retrieval evaluations in Table 1 then predict or retrieve exactly those grouped labels or metadata texts. Consequently, high scores on the internal test split largely certify that the contrastive objective was optimized rather than providing independent evidence that the representations generalize. This is a real but partial circularity: the OASIS experiments in Table 3 are out-of-distribution and subject-disjoint, so they give independent grounding for the generalization claim. No self-citation chain is load-bearing here; the cited methods (SupCon, CLIP, OpenCLIP) are external. The scan-level split concern raised in the skeptic note is a data-leakage/validity issue rather than a circularity and is not scored as such. Overall, the paper is not wholly circular, but its headline in-house numbers should be read as verification of the training signal rather than as external validation.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on two domain assumptions about the reliability of DICOM metadata and the validity of the discretization grouping, plus the standard assumptions of contrastive representation learning. There are no fitted parameters in a mathematical derivation; the main free choices are the discretization grid and TI binning, which are hand-picked and shown to affect performance.

free parameters (2)

TE-TR discretization grid size = 20x20 in main experiments; 5x5 gave best retrieval in ablation
Chosen by hand in Section 2 (label construction); Table 2 shows retrieval scores change by up to 20 points across grid sizes, so this choice materially affects the reported numbers.
TI binning scheme = separate sparse bins for TI; >85% of scans have no inversion pulse
Chosen by hand in Section 2 to handle the long-tailed TI distribution; it directly shapes which scans are treated as positives in Eq. (1).

assumptions (3)

domain assumption DICOM metadata fields accurately and comprehensively determine MR image contrast.
This is the supervision source for the entire method; it is explicitly acknowledged as an assumption in the Discussion and is the load-bearing premise for the grouping strategy.
domain assumption Scans sharing the same grouped metadata label have equivalent visual contrast.
Required for the SupCon positive sets P(i) in Eq. (1) to be correct; if discretization merges distinct contrasts, the supervision is noisy or wrong.
standard math The Supervised Contrastive Loss yields useful representations under standard conditions.
The paper relies on the known properties of InfoNCE/SupCon (refs 7, 12) without extending the theory; this is standard background in contrastive learning.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MR-CLIP: Efficient Metadata-Guided Learning of MRI Contrast Representations." pith.science (2026). https://pith.science/paper/5HFW5SXO

@misc{pith2026250700043,
  author       = {Pith},
  title        = {Pith review of: MR-CLIP: Efficient Metadata-Guided Learning of MRI Contrast Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5HFW5SXO}},
  note         = {Machine review of arXiv:2507.00043}
}

read the original abstract

Accurate interpretation of Magnetic Resonance Imaging scans in clinical systems is based on a precise understanding of image contrast. This contrast is primarily governed by acquisition parameters, such as echo time and repetition time, which are stored in the DICOM metadata. To simplify contrast identification, broad labels such as T1-weighted or T2-weighted are commonly used, but these offer only a coarse approximation of the underlying acquisition settings. In many real-world datasets, such labels are entirely missing, leaving raw acquisition parameters as the only indicators of contrast. Adding to this challenge, the available metadata is often incomplete, noisy, or inconsistent. The lack of reliable and standardized metadata complicates tasks such as image interpretation, retrieval, and integration into clinical workflows. Furthermore, robust contrast-aware representations are essential to enable more advanced clinical applications, such as achieving modality-invariant representations and data harmonization. To address these challenges, we propose MR-CLIP, a multimodal contrastive learning framework that aligns MR images with their DICOM metadata to learn contrast-aware representations, without relying on manual labels. Trained on a diverse clinical dataset that spans various scanners and protocols, MR-CLIP captures contrast variations across acquisitions and within scans, enabling anatomy-invariant representations. We demonstrate its effectiveness in cross-modal retrieval and contrast classification, highlighting its scalability and potential for further clinical applications. The code and weights are publicly available at https://github.com/myigitavci/MR-CLIP.

Figures

Figures reproduced from arXiv: 2507.00043 by the authors.

**Figure 1.** MR-CLIP leverages slice-level image and metadata pairs across multiple acquisitions and slices. An image encoder processes slices, while a text encoder embeds corresponding structured DICOM metadata, with MR-CLIP contrastively learning to align image and metadata representations. Representations of slices of similar acquisitions (Acq. 1 and Acq. 3) are pulled together, while those of distinct contrasts (Acq. 2) ar… view at source ↗

**Figure 2.** Overview of metadata distribution in our dataset. (a) Categorical tags including scanner, plane, field strength, and sequence information and flip angle. (b) Numerical distribution of echo and repetition times, color-coded by inversion time. is excluded from label construction due to site-dependent variability; instead, it is used only as an auxiliary input. Unlike traditional CLIP, which uses one-to-one image-text … view at source ↗

**Figure 3.** Error rates across DICOM tags based on linear probe classification results. approximate deviations of 14.4 ms in TE and 1145 ms in TR (based on a 20×20 TE–TR grid). This suggests that while predictions often land near the correct value, they are not always in the exact same bin, likely due to the semantic ambiguity introduced by discretization. In contrast, clearly defined tags such as Field Strength show near-zero … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 16 canonical work pages

[1]

Chung, H., Lee, D., Wu, Z., Kim, B.H., Bouman, K.L., Ye, J.C.: ContextMRI: EnhancingcompressedsensingMRIthroughmetadataconditioning.arXivpreprint arXiv:2501.04284 (2025)

arXiv 2025
[2]

In: Proceedings of the Third Conference on Medical Imaging with Deep Learning

Du, R., Vardhanabhuti, V.: 3D-RADNet: Extracting labels from DICOM meta- data for training general medical domain deep 3D convolution neural networks. In: Proceedings of the Third Conference on Medical Imaging with Deep Learning. vol. 121, pp. 174–192. PMLR (2020)

work page 2020
[3]

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

Gauriau, R., Bridge, C., Chen, L., Glocker, B., Hajnal, J., Rueckert, D., Bai, W.: Using DICOM metadata for radiological image series categorization: A feasibility study on large clinical brain MRI datasets. Journal of Digital Imaging33(3), 747– 762 (2020)

work page 2020
[4]

Medical Image Analysis97, 103296 (2024)

Holland,R.,Leingang,O.,Bogunović,H.,Riedl,S.,Fritsche,L.,Prevost,T.,Scholl, H.P., Schmidt-Erfurth, U., Sivaprasad, S., Lotery, A.J., Rueckert, D., Menten, M.J.: Metadata-enhanced contrastive learning from retinal optical coherence to- mography images. Medical Image Analysis97, 103296 (2024)

work page 2024
[5]

NeuroImage260, 119474 (2022)

Hoopes, A., Mora, J.S., Dalca, A.V., Fischl, B., Hoffmann, M.: Synthstrip: skull- stripping for any brain image. NeuroImage260, 119474 (2022)

2022
[6]

Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021)

work page 2021
[7]

In: Proceedings of the 34th International Conference on Neural Information Processing Systems

Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: Proceedings of the 34th International Conference on Neural Information Processing Systems. NIPS ’20, Curran Associates Inc., Red Hook, NY, USA (2020)

work page 2020
[8]

medRxiv (2019)

LaMontagne, P.J., Benzinger, T.L., Morris, J.C., Keefe, S., Hornbeck, R., Xiong, C., Grant, E., Hassenstab, J., Moulder, K., Vlassenko, A.G., Raichle, M.E., Cruchaga, C., Marcus, D.: OASIS-3: Longitudinal neuroimaging, clinical, and cog- nitive dataset for normal aging and Alzheimer disease. medRxiv (2019)

work page 2019

Show all 20 references

[9]

Frontiers in Neuroinformatics15, 622951 (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., ONDRI Investiga- tors, Strother, S.C.: Magnetic Resonance I...

2021
[10]

Academic Radiology (2022)

McDaniel, J.W., Moore, A.Z., Mareci, T.H., Price, S.L., Conklin, D.J., Wagle, N., Pinho, M.C., Landman, B.A.: Improving the automatic classification of brain MRI acquisition contrast with machine learning. Academic Radiology (2022)

2022
[11]

https://www.dicomstandard.org/ (May 2025), nEMA PS3 / ISO 12052 10 M.Y

National Electrical Manufacturers Association: Digital imaging and communi- cations in medicine (DICOM) standard. https://www.dicomstandard.org/ (May 2025), nEMA PS3 / ISO 12052 10 M.Y. Avci et al

2025
[12]

arXiv preprint arXiv:1807.03748 (2019)

van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2019)

2019 arXiv
[13]

In: Information Processing in Medical Imaging (IPMI)

Ouyang, J., Adeli, E., Pohl, K.M., Zhao, Q., Zaharchuk, G.: Representation dis- entanglement for multi-modal brain MRI analysis. In: Information Processing in Medical Imaging (IPMI). Lecture Notes in Computer Science, vol. 12729, pp. 321–
[14]

In: Proceedings of the International Conference on Machine Learning (2021)

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning trans- ferable visual models from natural language supervision. In: Proceedings of the International Conference on Machin...

2021
[15]

Neuroinformatics22, 297–315 (2024)

Sinha, H., Raamana, P.R.: Solving the pervasive problem of protocol non- compliance in MRI using an open-source tool mrQA. Neuroinformatics22, 297–315 (2024)

2024
[16]

In: International Conference on Medical Image Comput- ing and Computer-Assisted Intervention (2022)

Wang, H., Liu, K., Ng, N., et al.: MedCLIP: Contrastive learning from unpaired medical images and text. In: International Conference on Medical Image Comput- ing and Computer-Assisted Intervention (2022)

2022
[17]

arXiv preprint arXiv:2409.16818 (2024)

Wang, Y., Xiong, H., Sun, K., Bai, S., Dai, L., Ding, Z., Liu, J., Wang, Q., Liu, Q., Shen, D.: Towards general text-guided image synthesis for customized multimodal brain MRI generation. arXiv preprint arXiv:2409.16818 (2024)

2024 arXiv
[18]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)

Wu, J., Zhang, Y., Xie, Y., Ma, T., et al.: PMC-CLIP: Contrastive vision-language pretraining on biomedical literature. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)

2023
[19]

NEJM AI2(1) (2024)

Zhang, S., Xu, Y., Usuyama, N., Xu, H., Bagga, J., Tinn, R., Preston, S., Rao, R., Wei, M., Valluri, N., Wong, C., Tupini, A., Wang, Y., Mazzola, M., Shukla, S., Liden, L., Gao, J., Crabtree, A., Piening, B., Bifulco, C., Lungren, M.P., Naumann, T., Wang, S., Poon, H.: A multi...

2024
[20]

Comput- erized Medical Imaging and Graphics109, 102285 (2023)

Zuo, L., Liu, Y., Xue, Y., Dewey, B.E., Remedios, S.W., Hays, S.P., Bilgel, M., Mowry, E.M., Newsome, S.D., Calabresi, P.A., Resnick, S.M., Prince, J.L., Carass, A.: HACA3: A unified approach for multi-site MR image harmonization. Comput- erized Medical Imaging and Graphics109...

2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Chung, H., Lee, D., Wu, Z., Kim, B.H., Bouman, K.L., Ye, J.C.: ContextMRI: EnhancingcompressedsensingMRIthroughmetadataconditioning.arXivpreprint arXiv:2501.04284 (2025)

arXiv 2025

[2] [2]

In: Proceedings of the Third Conference on Medical Imaging with Deep Learning

Du, R., Vardhanabhuti, V.: 3D-RADNet: Extracting labels from DICOM meta- data for training general medical domain deep 3D convolution neural networks. In: Proceedings of the Third Conference on Medical Imaging with Deep Learning. vol. 121, pp. 174–192. PMLR (2020)

work page 2020

[3] [3]

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

Gauriau, R., Bridge, C., Chen, L., Glocker, B., Hajnal, J., Rueckert, D., Bai, W.: Using DICOM metadata for radiological image series categorization: A feasibility study on large clinical brain MRI datasets. Journal of Digital Imaging33(3), 747– 762 (2020)

work page 2020

[4] [4]

Medical Image Analysis97, 103296 (2024)

Holland,R.,Leingang,O.,Bogunović,H.,Riedl,S.,Fritsche,L.,Prevost,T.,Scholl, H.P., Schmidt-Erfurth, U., Sivaprasad, S., Lotery, A.J., Rueckert, D., Menten, M.J.: Metadata-enhanced contrastive learning from retinal optical coherence to- mography images. Medical Image Analysis97, 103296 (2024)

work page 2024

[5] [5]

NeuroImage260, 119474 (2022)

Hoopes, A., Mora, J.S., Dalca, A.V., Fischl, B., Hoffmann, M.: Synthstrip: skull- stripping for any brain image. NeuroImage260, 119474 (2022)

2022

[6] [6]

Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021)

work page 2021

[7] [7]

In: Proceedings of the 34th International Conference on Neural Information Processing Systems

Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: Proceedings of the 34th International Conference on Neural Information Processing Systems. NIPS ’20, Curran Associates Inc., Red Hook, NY, USA (2020)

work page 2020

[8] [8]

medRxiv (2019)

LaMontagne, P.J., Benzinger, T.L., Morris, J.C., Keefe, S., Hornbeck, R., Xiong, C., Grant, E., Hassenstab, J., Moulder, K., Vlassenko, A.G., Raichle, M.E., Cruchaga, C., Marcus, D.: OASIS-3: Longitudinal neuroimaging, clinical, and cog- nitive dataset for normal aging and Alzheimer disease. medRxiv (2019)

work page 2019