Pith. sign in

REVIEW 4 major objections 5 minor 44 references

Improving Factuality of 3D Brain MRI Report Generation with Paired Image-domain Retrieval and Text-domain Augmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Retrieving similar 3D DWI/ADC volumes and grounding an LLM on their paired reports yields more factually accurate ischemic-stroke radiology reports than direct image-to-text generation with GPT-4o or LLaVA-Med.

desk verdict A genuinely useful retrieval-based framing for 3D MRI report generation, with a confounded baseline comparison and metric circularity that leave the central claim under-supported. read the letter →

arxiv 2411.15490 v3 pith:TQL23TL2 submitted 2024-11-23 cs.CV cs.LGeess.IV

Junhyeok Lee , Yujin Oh , Dahyoun Lee , Hyon Keun Joh , Minchul Kim , Chul-Ho Sohn , Sung Hyun Baik , Cheol Kyu Jung
show 4 more authors
Jung Hyun Park Kyu Sung Choi Byung-Hoon Kim Jong Chul Ye
This is my paper · ORCID
classification cs.CVcs.LGeess.IV
keywords retrieval-augmentedgeneration3DbrainMRIDWI/ADCacuteischemicstrokeradiologyreportmaskedautoencodervisiontransformerterritoryclassification
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

PIRTA aims to establish that factually reliable radiology reports for 3D brain MRI can be generated by retrieving clinically similar DWI/ADC volumes and reusing their paired clinician-authored reports, rather than by learning to map images directly to text. The paper's argument is that the hard part of image-to-text generation is explicit cross-modal alignment, so it avoids that step entirely: only a 3D vision encoder is trained, and the retrieved reports supply the clinical text that grounds an LLM. On internal and external datasets the framework reaches image-retrieval mAP@1 of 94.04% internally and 71.21% and 70.87% on two external cohorts, with ischemic-territory accuracy in generated reports consistently higher than GPT-4o and LLaVA-Med given 2D slices. If true, this means report factuality for 3D medical imaging can be improved with retrieval infrastructure and paired historical reports instead of larger multimodal alignment models.

What carries the argument

The central object is PIRTA, a paired image-domain retrieval and text-domain augmentation pipeline. Its engine is a 3D Vision Transformer image encoder: stage one pretrains it as a masked autoencoder on two-channel DWI+ADC volumes, and stage two fine-tunes it with a linear classifier over four classes (anterior circulation, deep gray matter, posterior circulation, normal). During inference the encoder is frozen, cosine similarity over its embeddings retrieves the top m=5 database images, and their paired structured findings, similarity scores, and patient registry data are fed as instruction context to a LoRA-tuned LLaMA-3-8B-Instruct that outputs clinical presentation, NIHSS, past history, findings, and impression. This machinery carries the argument because it replaces cross-modal alignment with image-only representation learning plus reuse of verified text.

What would settle it

Compare PIRTA's generated-report territory accuracy with a retrieval-only oracle that outputs the territory label of the top-1 retrieved image's paired report; if generated-report accuracy never exceeds the oracle, the LLM adds no independent factual information and report factuality is exactly retrieval accuracy. A direct test would use rare sites such as isolated amygdala infarction, where similarity scores drop, and check whether every generation error coincides with a retrieval error.

Watch

Extended reading notes

Core claim

The central discovery is that the cross-modal mapping problem in radiology-report generation can be recast as an in-domain retrieval problem. Instead of learning a text encoder and aligning image and text representations in a shared space, PIRTA trains only a 3D ViT image encoder and retrieves the database image that minimizes the cosine distance between query and database embeddings; the paired report of each retrieved image then augments LLM generation. The paper shows that a 3D ViT pretrained with masked autoencoding on over 38,000 unlabeled DWI/ADC volumes and fine-tuned to classify four ischemic territories produces retrieval that transfers to external institutions and to the public ISLES benchmark, and that grounding LLaMA-3-8B-Instruct on the top-5 retrieved reports yields higher ischemic-territory accuracy than direct generation by GPT-4o or LLaVA-Med. The paper frames this as a complexity reduction from learning $C(\mathcal{H}_{\text{image}}) + C(\mathcal{H}_{\text{text}})$ to learning $C(\mathcal{H}_{\text{image}})$ alone.

Load-bearing premise

Everything depends on the assumption that images the encoder finds similar truly share the same ischemic findings, so the retrieved paired reports are the right text to ground the generated report.

Editorial extensions

If this is right

  • Direct image-to-text alignment is not required for factual 3D MRI report generation; a paired image-report database can substitute for a learned cross-modal mapping.
  • Pretraining scale is a first-order driver: adding 38,532 unlabeled UK Biobank volumes raised retrieval mAP@1 from 80.73% to 94.04% internally, from 56.72% to 71.21% on BRMH, and from 38.83% to 70.87% on ISLES.
  • Retrieval-grounded generation transfers across institutions and scanner vendors, so a system built this way can be externally validated without retraining the text side.
  • The output includes the top-5 retrieved reports and their similarity scores, giving clinicians auditable evidence for each generated finding.

Reading between the lines

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

  • Implicit upper bound: generated-report factuality cannot exceed retrieval quality, so the framework's benefit over direct generation should be measured against an oracle that simply copies the top-1 retrieved report's territory label.
  • The printed similarity scores could serve as an uncertainty flag: low top-5 scores (the failure case shows 0.53 to 0.46) mark cases that should route to human review rather than autonomous sign-off.
  • The same recipe—masked-autoencoder pretraining, territory fine-tuning, cosine retrieval, paired-report grounding—should transfer to other 3D modalities such as CT or to pathology slides whenever a paired verified-report database exists.
  • Because the encoder is fine-tuned on the same four territory labels used to judge retrieval relevance, an independent evaluation should test retrieval against richer clinical ground truth such as laterality, lesion size, or multi-territory involvement.
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.

Referee Report

4 major / 5 minor

Summary. The paper proposes PIRTA, a retrieval-augmented generation framework for 3D DWI/ADC brain MRI reports. A 3D ViT is pretrained with masked autoencoding on a large unlabeled dataset, supervised fine-tuned to classify four ischemic-territory/normal labels, and then used to retrieve top-k similar images from a paired image-report database. The paired clinician-authored reports are provided as in-context text to a LoRA-tuned LLaMA-3-8B-Instruct, which generates a structured radiology report. Experiments on an internal set (SNUH+SNUBH) and external sets (BRMH, ISLES) report high retrieval mAP/Acc@k and higher ischemic-territory Acc@1 than GPT-4o and LLaVA-Med baselines, which receive only a single 2D slice and one in-context example. The paper argues that by avoiding explicit image-text alignment and instead retrieving paired reports, PIRTA improves factual accuracy of generated reports.

Significance. If the central claim holds, the work is a useful step toward clinically grounded report generation from volumetric brain MRI, with a practical recipe (MAE pretraining + supervised fine-tuning on a small label set + nearest-neighbor report retrieval + LLM augmentation) that avoids expensive cross-modal alignment. Strengths of the manuscript include the use of two external datasets (BRMH and ISLES), a large unlabeled pretraining corpus (UK Biobank), and the release of source code. The retrieval results in Table 4, especially the improvement from large-scale MAE pretraining, are internally consistent and are a solid empirical contribution. However, the paper's headline claim—that retrieval-grounded generation improves report factuality over direct image-to-text models—is currently under-supported because the comparison to GPT-4o and LLaVA-Med changes several variables at once, and the retrieval evaluation is tied to the same four labels used to fine-tune the encoder.

major comments (4)
  1. [Section 3.1.2 and Figure 3] The central empirical comparison is confounded. PIRTA receives full 3D DWI+ADC volumes, five retrieved clinician-authored reports, a LoRA-tuned LLaMA-3 that has been instruction-tuned exactly on the structured template, and additional clinical context fields; GPT-4o and LLaVA-Med receive a single 2D axial slice, one in-context example of the desired format, and no retrieval or clinical context. The reported advantage in ischemic-territory Acc@1 could therefore be due to the 3D input, the LLM instruction-tuning, the extra clinical variables, or the retrieved reports rather than to retrieval-grounded generation per se. The manuscript should include an ablation that holds the 3D encoder and the LoRA-tuned LLM fixed and toggles only the retrieved text (e.g., PIRTA without retrieved reports, PIRTA with randomly sampled retrieved reports, or both). Without such an ablation, the title claim that retrieval-grounded generation improves factuality is not established.
  2. [Section 2.2, Eqs. (8)-(9)] The complexity argument does not support the factual-superiority claim. Equation (9) reduces to C(H_image) + C(H_text) >= C(H_image), which is trivially true because C(H_text) is non-negative. This bound concerns the size of a hypothesis space and says nothing about whether retrieved reports will be clinically relevant or whether generated reports will be factual. The discussion in Section 2.2 ("The retrieval of relevant text information from the paired database ensures factual accuracy") asserts rather than demonstrates the link between retrieval relevance and report factuality. The paper needs a quantitative analysis connecting retrieval quality (e.g., top-5 agreement or similarity score) to generation accuracy, or a separate experiment that corrupts the retrieved context to show its causal effect.
  3. [Table 4 and Section 3.1.1] The retrieval evaluation is circular with respect to the generation target. The image encoder is supervised fine-tuned to classify exactly the four categories (normal, anterior, deep gray, posterior) that define relevance in Table 4, and the generated-report evaluation in Figure 3 uses the same four-category ischemic-territory accuracy. Thus the high mAP@1 and Acc@1 values in Table 4 demonstrate that the encoder can recognize the four training labels on internal and external data; they do not independently establish that the top-k retrieved reports contain the clinically relevant findings for the query image. The authors should either evaluate retrieval with a clinician-defined relevance judgment that goes beyond the four territory labels, or explicitly acknowledge that Table 4 measures a recognition performance that is partially enforced by the SFT stage.
  4. [Section 3.2.2 / Table 5 and Discussion] The paper's own failure case shows that when top-5 similarity scores are low (0.53 to 0.46), the generated report is wrong. This is an honest and informative observation, but the manuscript does not quantify how often such low-confidence retrievals occur or how generation accuracy varies with retrieval similarity. Given the paper's central thesis that factuality is inherited from retrieval, the authors should report generation Acc@1 stratified by retrieval similarity (e.g., bins of top-1 or top-5 similarity) or by retrieval success/failure on the internal and external test sets. This would also provide a clinically useful operating characteristic for PIRTA and would prevent the reader from having to rely on a single qualitative example.
minor comments (5)
  1. [Section 2.3.2] The handling of the MAE mask ratio is unclear: the text says "a random mask is applied to remove (1-q) of the patches" and later sets "mask ratio q of 0.25," which would mean 75% of patches are removed; typically a 0.25 mask ratio means 25% are removed. Please clarify the intended keep/remove proportion.
  2. [Section 3.1.2] No details are provided on how the "center of the stroke lesion" slice was identified for GPT-4o and LLaVA-Med; if lesion localization was used to choose the input slice, the baseline setup should be described precisely, and the authors should discuss whether this favors or disadvantages the baselines.
  3. [Figure 3 and Table 6] The figure and tables report point estimates without confidence intervals, error bars, or significance tests. Given the small numbers in some external classes (e.g., ISLES deep-gray n=52 in Table 2), the authors should report variance or at least a statistical comparison for the main Acc@1 differences.
  4. [Table 2] The text in Section 2.3.1 says the ISLES 2022 dataset includes 211 subjects, but Table 2 lists 206 ISLES test samples; please reconcile this discrepancy.
  5. [References] Several references are incomplete or formatted inconsistently (e.g., Ram et al., Borgeaud et al., and Karpukhin et al. contain ellipses in the author lists, and some entries have inconsistent venue naming). Please ensure all references are complete and uniform.

Circularity Check

1 steps flagged · score 6.0 of 10

The report-factuality metric is the same four-class SFT target used to train the retrieval encoder, and the generated territory is copied from retrieval, so the main result partially reduces to the classifier's training objective.

  1. fitted input called prediction [Section 2.3.2 and Section 2.3.3; Figure 3 and Table 4]
    "In the second stage, the image encoder is fine-tuned to classify the stroke lesion into one of the four classes, i.e. anterior, deep gray, posterior, and normal. ... To further evaluate the performance of text-domain augmentation, we conducted a quantitative evaluation of the Acc@1 of ischemic territory information extracted from the generated reports, comparing these results with those obtained from image-domain retrieval."

    The four SFT classes are the same labels used to define retrieval relevance in Table 4 and to score report factuality in Figure 3. The LLM receives no image input; its only image-derived information is the top-5 retrieved reports, so the generated territory is inherited from the retrieval result, which is a k-NN decision in the SFT embedding space. Thus the reported improvement in ischemic-territory accuracy restates the fine-tuned classifier's accuracy under the name of report generation rather than independently validating text-domain augmentation. Table 5 confirms the inheritance: when retrieval says deep gray, the generated report says deep gray, while the ground truth is anterior circulation.

full rationale

The central reported result is that PIRTA improves ischemic-territory accuracy over direct image-to-text baselines. I found one partial circularity: the image encoder is supervised fine-tuned on exactly the four classes (normal, anterior, deep gray, posterior) that define both retrieval relevance in Table 4 and the report-factuality metric in Figure 3. Because the LLM receives the query image only through the top-5 retrieved reports, the territory in the generated report is inherited from the retrieved reports; the failure case in Table 5 shows this inheritance explicitly, with retrieval saying deep gray, generation saying deep gray, and the ground truth being anterior. Thus the prediction of report territory is, by construction, the SFT classifier's k-NN output rather than an independent text-generation result. The external BRMH and ISLES results show that the trained encoder generalizes to held-out institutions and benchmarks, but they do not break the reduction because the same label scheme is both the training target and the evaluation metric. No load-bearing self-citation or imported uniqueness theorem appears; the cited prior work by the same group is not load-bearing. The comparison to GPT-4o and LLaVA-Med changes multiple variables simultaneously, including 2D-slice input versus 3D volumes, one-shot prompting versus LoRA fine-tuning, and no retrieval versus retrieval; that is a real experimental confound, but it is a correctness concern rather than a circularity. Overall, the central factuality claim is partially circular because the measured outcome reduces to the supervised classification objective used to build the retrieval space, so the score is 6.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim depends on five stated assumptions: transferability of MAE pretraining to clinical DWI/ADC, clinical validity of cosine similarity in the fine-tuned embedding space, sufficiency of territory accuracy as a factuality surrogate, fairness of the 2D baseline setup, and an unproven additive complexity measure in Section 2.2. No new physical entities are introduced. The numeric hyperparameters (m=5, mask ratio 0.25, taxonomy) are hand-chosen and affect the results.

free parameters (3)
  • m = 5 (number of retrieved reports) = 5
    The number of retrieved images/reports provided to the LLM is a hand-chosen hyperparameter that directly determines the context available for generation and the failure mode at low similarity.
  • MAE mask ratio q = 0.25 = 0.25
    Mask ratio for SSL pretraining is chosen by hand; retrieval performance depends on the learned representations, and no sensitivity analysis is provided.
  • Structured findings taxonomy (2 classes, 5 infarction types, 3 territories) = per-neuroradiologist labels
    The label space used for SFT and for the evaluation metric is a hand-designed simplification of real radiology reports; the central claim is measured entirely within this taxonomy.
assumptions (5)
  • domain assumption UK Biobank DWI/ADC MAE pretraining transfers to clinical multi-vendor DWI/ADC volumes.
    The image encoder is pretrained on unlabeled UK Biobank data and then fine-tuned on in-house data; the retrieval and report-generation results depend on this transfer (Section 2.3.2).
  • domain assumption Cosine similarity in the SFT-tuned embedding space ranks images by clinically relevant report content.
    Section 2.3.2 retrieves by cosine similarity of fine-tuned ViT features; the qualitative failure in Table 5 shows this similarity can fail for rare anatomies such as isolated amygdala infarction.
  • domain assumption Ischemic-territory accuracy is a sufficient surrogate for report factuality.
    Section 3.1.2 uses territory Acc@1 as the primary metric for factuality; class and infarction type and overall report correctness are not quantitatively evaluated, so the factuality claim is narrower than stated in the title.
  • domain assumption A single 2D axial slice (middle or lesion center) is a fair input for GPT-4o and LLaVA-Med baselines.
    Section 3.1.2 compares 2D-slice baselines against full-3D PIRTA, assuming the slice preserves the territory information needed for a fair comparison; no baseline gets the same 3D volume or retrieved context.
  • ad hoc to paper The complexity measure C in Eq. (5) is additive over hypothesis spaces.
    Eq. (5) C(H) = C(H_image) + C(H_text) is introduced without a formal definition or proof, and the conclusion in Eq. (9) is a restatement of non-negativity; this is used to argue PIRTA simplifies the learning task.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Factuality of 3D Brain MRI Report Generation with Paired Image-domain Retrieval and Text-domain Augmentation." pith.science (2026). https://pith.science/paper/TQL23TL2

@misc{pith2026241115490,
  author       = {Pith},
  title        = {Pith review of: Improving Factuality of 3D Brain MRI Report Generation with Paired Image-domain Retrieval and Text-domain Augmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TQL23TL2}},
  note         = {Machine review of arXiv:2411.15490}
}
read the original abstract

Acute ischemic stroke (AIS) requires time-critical decision-making, where inaccurate interpretation of neuroimaging findings can lead to irreversible disability. Diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) maps from magnetic resonance imaging (MRI) are central to detecting acute infarction, yet generating factually reliable radiology reports directly from 3D MRI remains challenging due to the difficulty of learning robust cross-modal alignments between volumetric images and clinical text. We propose paired image-domain retrieval and text-domain augmentation (PIRTA), a retrieval-augmented generation framework that improves report factuality by avoiding explicit image-text alignment. PIRTA retrieves clinically similar 3D DWI/ADC volumes using a pretrained 3D vision encoder and leverages their paired clinician-authored reports to ground large language model (LLM)-based report generation. Experiments on multi-institutional in-house data, a held-out external privacy-preserving cohort, and the public ISLES benchmark demonstrate that PIRTA achieves strong image-domain retrieval performance and consistently improves ischemic-territory accuracy, a clinically grounded surrogate for report factuality, compared to direct image-to-text baselines. These results indicate that retrieval-grounded generation provides a scalable and reliable paradigm for producing factually consistent radiology reports from complex 3D brain MRI. Source code is available at https://github.com/jhlee0619/PIRTA.

Figures

Figures reproduced from arXiv: 2411.15490 by the authors.

Figure 1
Figure 1. Graphical illustration of the proposed method for image-to-text retrieval. a) Conventional methods require training an image encoder [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Schematic illustration of the proposed method. a) Overview of training the 3D MRI image encoder. In the first stage, the 3D MRI input [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Quantitative performance of the proposed method compared to baseline. SNUH [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Retrieval results on three classes. five scans are retrieved for each query. The first column displays the query image, while the following [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: The 768-dimensional feature vectors derived image encoder are projected into a 2D manifold using UMAP. Each stroke lesion type is [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 34 canonical work pages

  1. [1]

    , author Jenkinson, M

    author Alfaro-Almagro, F. , author Jenkinson, M. , author Bangerter, N.K. , author Andersson, J.L. , author Griffanti, L. , author Douaud, G. , author Sotiropoulos, S.N. , author Jbabdi, S. , author Hernandez-Fernandez, M. , author Vallee, E. , et al., year 2018 . title Image processing and quality control for the first 10,000 brain imaging datasets from ...

    work page 2018

  2. [2]

    , author Coutinho, A.M.N

    author Alves, I.S. , author Coutinho, A.M.N. , author Vieira, A.P.F. , author Rocha, B.P. , author Passos, U.L. , author Gon c alves, V.T. , author Silva, P.D.S. , author Zhan, M.X. , author Pinho, P.C. , author Delgado, D.S. , author Docema, M.F.L. , author Lee, H.W. , author Policeni, B.A. , author Leite, C.C. , author Martin, M.G.M. , author Amancio, C...

    work page 2022

  3. [3]

    , author Cook, M.J

    author Beh, S.M.J. , author Cook, M.J. , author D'Souza, W.J. , year 2016 . title Isolated amygdala enlargement in temporal lobe epilepsy: A systematic review . journal Epilepsy Behav. volume 60 , pages 33--41

    work page 2016

  4. [4]

    , author Hudson, D.A

    author Bommasani, R. , author Hudson, D.A. , author Adeli, E. , author Altman, R. , author Arora, S. , author von Arx, S. , author Bernstein, M.S. , author Bohg, J. , author Bosselut, A. , author Brunskill, E. , et al., year 2021 . title On the opportunities and risks of foundation models . journal arXiv preprint arXiv:2108.07258

    arXiv 2021

  5. [5]

    , author Mensch, A

    author Borgeaud, S. , author Mensch, A. , author Hoffmann, J. , author Cai, T. , author Rutherford, E. , author Millican, K. , author ... , author Irving, G. , year 2022 . title Improving language models by retrieving from trillions of tokens , in: booktitle Proceedings of the 39th International Conference on Machine Learning (ICML) , publisher PMLR . pp....

    work page 2022

  6. [6]

    , author De Silva, D.A

    author Campbell, B.C. , author De Silva, D.A. , author Macleod, M.R. , author Coutts, S.B. , author Schwamm, L.H. , author Davis, S.M. , author Donnan, G.A. , year 2019 . title Ischaemic stroke . journal Nature reviews Disease primers volume 5 , pages 70

    work page 2019

  7. [7]

    , author Touvron, H

    author Caron, M. , author Touvron, H. , author Misra, I. , author J \'e gou, H. , author Mairal, J. , author Bojanowski, P. , author Joulin, A. , year 2021 . title Emerging properties in self-supervised vision transformers , in: booktitle Proceedings of the IEEE/CVF international conference on computer vision , pp. pages 9650--9660

    2021

  8. [8]

    , author Koska, I.O

    author Cetinoglu, Y.K. , author Koska, I.O. , author Uluc, M.E. , author Gelal, M.F. , year 2021 . title Detection and vascular territorial classification of stroke on diffusion-weighted mri by deep learning . journal European Journal of Radiology volume 145 , pages 110050

    work page 2021

Show all 44 references
  1. [9]

    , author Beyer, L

    author Dosovitskiy, A. , author Beyer, L. , author Kolesnikov, A. , author Weissenborn, D. , author Zhai, X. , author Unterthiner, T. , author Dehghani, M. , author Minderer, M. , author Heigold, G. , author Gelly, S. , et al., year 2020 . title An image is worth 16x16 words: ...

    2020 arXiv

  2. [10]

    , author Brainin, M

    author Feigin, V.L. , author Brainin, M. , author Norrving, B. , author Martins, S. , author Sacco, R.L. , author Hacke, W. , author Fisher, M. , author Pandian, J. , author Lindsay, P. , year 2022 . title World stroke organization (wso): global stroke fact sheet 2022 . journa...

    2022

  3. [11]

    , author Krishnamurthi, R.V

    author Feigin, V.L. , author Krishnamurthi, R.V. , author Parmar, P. , author Norrving, B. , author Mensah, G.A. , author Bennett, D.A. , author Barker-Collo, S. , author Moran, A.E. , author Sacco, R.L. , author Truelsen, T. , author Davis, S. , author Pandian, J.D. , author ...

    2013

  4. [12]

    , author Chen, X

    author He, K. , author Chen, X. , author Xie, S. , author Li, Y. , author Doll \'a r, P. , author Girshick, R. , year 2022 . title Masked autoencoders are scalable vision learners , in: booktitle Proceedings of the IEEE/CVF conference on computer vision and pattern recognition...

    2022

  5. [13]

    , author Kim, Y.S

    author Heo, T.S. , author Kim, Y.S. , author Choi, J.M. , author Jeong, Y.S. , author Seo, S.Y. , author Lee, J.H. , author Jeon, J.P. , author Kim, C. , year 2020 . title Prediction of stroke outcome using natural language processing-based machine learning of radiology report...

    2020

  6. [14]

    , author de la Rosa, E

    author Hernandez Petzsche, M.R. , author de la Rosa, E. , author Hanning, U. , author Wiest, R. , author Valenzuela, W. , author Reyes, M. , author Meyer, M. , author Liew, S.L. , author Kofler, F. , author Ezhov, I. , et al., year 2022 . title Isles 2022: A multi-center magne...

    2022

  7. [15]

    , author Murina, E

    author Herzog, L. , author Murina, E. , author D \"u rr, O. , author Wegener, S. , author Sick, B. , year 2020 . title Integrating uncertainty in deep neural networks for mri based stroke analysis . journal Medical image analysis volume 65 , pages 101790

    2020

  8. [16]

    , author Parmar, C

    author Hosny, A. , author Parmar, C. , author Quackenbush, J. , author Schwartz, L.H. , author Aerts, H.J.W.L. , year 2018 . title Artificial intelligence in radiology . journal Nat. Rev. Cancer volume 18 , pages 500--510

    2018

  9. [17]

    , author Shen, Y

    author Hu, E.J. , author Shen, Y. , author Wallis, P. , author Allen-Zhu, Z. , author Li, Y. , author Wang, S. , author Wang, L. , author Chen, W. , year 2021 . title LoRA: Low-Rank Adaptation of Large Language Models . journal arXiv preprint arXiv:2106.09685

    2021 arXiv

  10. [18]

    , author Xu, W

    author Huo, F. , author Xu, W. , author Guo, J. , author Wang, H. , author Guo, S. , year 2024 . title C2kd: Bridging the modality gap for cross-modal knowledge distillation , in: booktitle Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR...

    2024

  11. [19]

    , author Grave, E

    author Izacard, G. , author Grave, E. , year 2021 . title Leveraging passage retrieval with generative models for open domain question answering , in: booktitle Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Vo...

    2021

  12. [20]

    , author Tian, Y

    author Jing, L. , author Tian, Y. , year 2020 . title Self-supervised visual feature learning with deep neural networks: A survey . journal IEEE transactions on pattern analysis and machine intelligence volume 43 , pages 4037--4058

    2020

  13. [21]

    , author Oguz, B

    author Karpukhin, V. , author Oguz, B. , author Min, S. , author Lewis, P. , author Wu, L. , author Edunov, S. , author ... , author Yih, W.t. , year 2020 . title Dense passage retrieval for open-domain question answering , in: booktitle Proceedings of the 2020 Conference on E...

    2020

  14. [22]

    , author Ellsworth, W

    author Ke, A. , author Ellsworth, W. , author Banerjee, O. , author Ng, A.Y. , author Rajpurkar, P. , year 2021 . title Chextransfer: performance and parameter efficiency of imagenet models for chest x-ray interpretation , in: booktitle Proceedings of the Conference on Health,...

    2021

  15. [23]

    , author Selver, A

    author Koska, I.O. , author Selver, A. , author Gelal, F. , author Uluc, M.E. , author C etino g lu, Y.K. , author Yurttutan, N. , author Serindere, M. , author Dicle, O. , year 2024 . title Deep learning classification of ischemic stroke territory on diffusion-weighted mri: A...

    2024

  16. [24]

    , author Liu, C.C

    author Lee, K.Y. , author Liu, C.C. , author Chen, D.Y.T. , author Weng, C.L. , author Chiu, H.W. , author Chiang, C.H. , year 2023 a. title Automatic detection and vascular territory classification of hyperacute staged ischemic stroke on diffusion weighted image using convolu...

    2023

  17. [25]

    , author Kim, W.J

    author Lee, S. , author Kim, W.J. , author Ye, J.C. , year 2023 b. title Llm itself can read and generate cxr images . journal arXiv preprint arXiv:2305.11490

    2023 arXiv

  18. [26]

    u ttler, H. , author Lewis, M. , author Yih, W.t. , author Rockt \

    author Lewis, P. , author Perez, E. , author Piktus, A. , author Petroni, F. , author Karpukhin, V. , author Goyal, N. , author K \"u ttler, H. , author Lewis, M. , author Yih, W.t. , author Rockt \"a schel, T. , et al., year 2020 . title Retrieval-augmented generation for kno...

    2020

  19. [27]

    , author Wong, C

    author Li, C. , author Wong, C. , author Zhang, S. , author Usuyama, N. , author Liu, H. , author Yang, J. , author Naumann, T. , author Poon, H. , author Gao, J. , year 2024 . title Llava-med: Training a large language-and-vision assistant for biomedicine in one day . journal...

    2024

  20. [28]

    , author Zhao, Y

    author Liu, C.F. , author Zhao, Y. , author Yedavalli, V. , author Leigh, R. , author Falcao, V. , author STIR and VISTA Imaging investigators , author Miller, M.I. , author Hillis, A.E. , author Faria, A.V. , year 2023 . title Automatic comprehensive radiological reports for ...

    2023

  21. [29]

    , author Li, S

    author Ma, W. , author Li, S. , author Cai, L. , author Kang, J. , year 2024 . title Learning modality knowledge alignment for cross-modality transfer . journal arXiv preprint arXiv:2406.18864

    2024 arXiv

  22. [30]

    , author Schwartz, K.M

    author McDonald, R.J. , author Schwartz, K.M. , author Eckel, L.J. , author Diehn, F.E. , author Hunt, C.H. , author Bartholmai, B.J. , author Erickson, B.J. , author Kallmes, D.F. , year 2015 . title The effects of changes in utilization and technological advancements of cros...

    2015

  23. [31]

    , author Banerjee, O

    author Moor, M. , author Banerjee, O. , author Abad, Z.S.H. , author Krumholz, H.M. , author Leskovec, J. , author Topol, E.J. , author Rajpurkar, P. , year 2023 . title Foundation models for generalist medical artificial intelligence . journal Nature volume 616 , pages 259--265

    2023

  24. [32]

    , author Thurner, P

    author Mouridsen, K. , author Thurner, P. , author Zaharchuk, G. , year 2020 . title Artificial intelligence applications in stroke . journal Stroke volume 51 , pages 2573--2579

    2020

  25. [33]

    , author Orfanoudaki, A

    author Ong, C.J. , author Orfanoudaki, A. , author Zhang, R. , author Caprasse, F.P.M. , author Hutch, M. , author Ma, L. , author Fard, D. , author Balogun, O. , author Miller, M.I. , author Minnig, M. , et al., year 2020 . title Machine learning and natural language processi...

    2020

  26. [34]

    , author Rabinstein, A.A

    author Powers, W.J. , author Rabinstein, A.A. , author Ackerson, T. , author Adeoye, O.M. , author Bambakidis, N.C. , author Becker, K. , author Biller, J. , author Brown, M. , author Demaerschalk, B.M. , author Hoh, B. , et al., year 2019 . title Guidelines for the early mana...

    2019

  27. [35]

    , author Eisenschlos, J

    author Ram, O. , author Eisenschlos, J. , author Ra, S. , author ... , year 2023 . title Inference-time intervention for knowledge integration in generative models , in: booktitle Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL) , p...

    2023

  28. [36]

    , author Menon, B.K

    author San Rom \'a n, L. , author Menon, B.K. , author Blasco, J. , author Hern \'a ndez-P \'e rez, M. , author D \'a valos, A. , author Majoie, C.B. , author Campbell, B.C. , author Guillemin, F. , author Lingsma, H. , author Anxionnat, R. , et al., year 2018 . title Imaging ...

    2018

  29. [37]

    , author Tu, T

    author Singhal, K. , author Tu, T. , author Gottweis, J. , author Sayres, R. , author Wulczyn, E. , author Hou, L. , author Clark, K. , author Pfohl, S. , author Cole-Lewis, H. , author Neal, D. , et al., year 2023 . title Towards expert-level medical question answering with l...

    2023 arXiv

  30. [38]

    , author Tasci, I

    author Tasci, B. , author Tasci, I. , year 2022 . title Deep feature extraction based brain image classification model using preprocessed images: Pdrnet . journal Biomedical Signal Processing and Control volume 78 , pages 103948

    2022

  31. [39]

    title Tissue plasminogen activator for acute ischemic stroke

    author The National Institute of Neurological Disorders and Stroke rt-PA Stroke Study Group , year 1995 . title Tissue plasminogen activator for acute ischemic stroke . journal New England Journal of Medicine volume 333 , pages 1581--1588

    1995

  32. [40]

    , author Smith, C

    author Wardlaw, J.M. , author Smith, C. , author Dichgans, M. , year 2019 . title Small vessel disease: mechanisms and clinical implications . journal Lancet Neurol. volume 18 , pages 684--696

    2019

  33. [41]

    , author Zhang, Z

    author Xie, Z. , author Zhang, Z. , author Cao, Y. , author Lin, Y. , author Bao, J. , author Yao, Z. , author Dai, Q. , author Hu, H. , year 2022 . title Simmim: A simple framework for masked image modeling , in: booktitle Proceedings of the IEEE/CVF Conference on Computer Vi...

    2022

  34. [42]

    , author Campbell, B.C

    author Xiong, Y. , author Campbell, B.C. , author Schwamm, L.H. , author Meng, X. , author Jin, A. , author Parsons, M.W. , author Fisher, M. , author Jiang, Y. , author Che, F. , author Wang, L. , author Zhou, L. , author Dai, H. , author Liu, X. , author Pan, Y. , author Dua...

    2024

  35. [43]

    , author Liu, H

    author Zhou, L. , author Liu, H. , author Bae, J. , author He, J. , author Samaras, D. , author Prasanna, P. , year 2023 . title Self pre-training with masked autoencoders for medical image classification and segmentation , in: booktitle 2023 IEEE 20th International Symposium ...

    2023

  36. [44]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.