Pith. sign in

REVIEW 5 major objections 6 minor 46 references

FFA Sora, video generation as fundus fluorescein angiography simulator

T0 review · 5 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read FFA Sora turns text reports into 21-frame synthetic fluorescein angiography videos and reports low patient re-identification.

desk verdict First text-to-video FFA simulator, but the split and evaluation don't support the accuracy or privacy claims. read the letter →

arxiv 2412.17346 v1 pith:GET33JBD submitted 2024-12-23 cs.CV cs.AI

Xinyuan Wu , Lili Wang , Ruoyu Chen , Bowen Liu , Weiyi Zhang , Xi Yang , Yifan Feng , Mingguang He
show 1 more author
Danli Shi
This is my paper · ORCID
classification cs.CVcs.AI
keywords fundusfluoresceinangiographytext-to-videogenerationdiffusiontransformerwaveletflowvariationalautoencoderprivacy-preservingdatasharingmedicaleducationretinalvasculardiseasesyntheticvideoevaluation
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

FFA Sora is a text-to-video model that takes a written fundus fluorescein angiography (FFA) report and turns it into a 21-frame synthetic FFA video showing retinal vascular structures and lesions described in the text. The paper's central claim is that the generated videos accurately simulate disease features such as dye leakage, microaneurysms, and non-perfusion areas, while producing low image-retrieval recall (average Recall@K = 0.073) that the authors interpret as strong privacy preservation against patient re-identification. The evidence comes from a retrospective dataset of 3,625 videos paired with 1,814 reports, objective quality metrics (FVD = 329.78, LPIPS = 0.48, VQAScore = 0.61), a BERTScore of 0.35 between reports regenerated from generated videos and the original prompts, and a human rating of 1.570 on a scale where 1 is best. If the claim holds, hospitals could share synthetic FFA sequences for medical education and for training diagnostic AI without releasing patient-identifiable video.

What carries the argument

The carrying mechanism is a two-stage latent video generator adapted from an open-source text-to-video implementation. The Wavelet-Flow Variational Autoencoder (WF-VAE) compresses FFA clips into latent tensors, using wavelet decomposition to route low-frequency image energy through a Main Energy Flow Pathway while converting convolutional layers to CausalConv3D so the model sees temporal structure. A Diffusion Transformer (DiT) then performs iterative denoising in that latent space, with self-attention for spatial and temporal coherence, cross-attention to text embeddings from a pretrained text encoder, and gating mechanisms that refine feature fusion. The decoder reconstructs 21-frame, 512-by-512 videos, and the pipeline includes a frame-standardization step that uses reverse-chronological selection or linear interpolation to fix every input to 21 frames.

What would settle it

Generate clips from prompts that name one lesion in one retinal quadrant and a specific phase, then have masked ophthalmologists or an automated detector localize that lesion in the generated frames; if localization accuracy is no better than chance, the claim that FFA Sora accurately simulates disease features from text fails.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that a diffusion transformer conditioned on text can synthesize FFA video sequences that carry the same clinically relevant features as real examinations—leakage, neovascularization, microaneurysms, capillary non-perfusion, and macular edema—across common retinal and choroidal diseases, and that the synthesized videos are hard to match back to their source videos in a retrieval test. The paper demonstrates this by building FFA Sora, applying it to a held-out test set, and reporting that the generated videos score well on standard generative metrics, achieve an acceptable text-alignment score, and receive an average visual quality rating of 1.570, close to the best possible score of 1.

Load-bearing premise

The load-bearing premise is that each text report is a specific and accurate description of its paired FFA video, even though 3,625 videos share only 1,814 reports and no patient-level or eye-level split is reported, so the text-to-video alignment and its evaluation rest on an ambiguous pairing.

Editorial extensions

If this is right

  • Text prompts describing diabetic retinopathy, vein occlusion, and choroidal diseases can be turned into synthetic FFA videos suitable for teaching trainees to recognize dynamic dye behavior.
  • The low average Recall@K of 0.073 implies that synthetic FFA videos could be shared across institutions with a much smaller risk of re-identifying the patient behind the original video.
  • Because a report-generation model can regenerate a textual report from a generated video with BERTScore 0.35 against the prompt, the model offers a way to create large, labeled, text-video training pairs for downstream diagnostic AI.
  • The same DiT-based architecture can generate FFA videos for less common conditions such as uveitis and retinitis pigmentosa, broadening the range of teachable cases beyond what a single hospital can collect.

Reading between the lines

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

  • A stronger fidelity test than BERTScore would be lesion localization: prompting for 'microaneurysms in the superior arcade' and asking a masked grader to draw the lesion should place it at the specified location; the paper does not report location-level correspondence.
  • The low Recall@K could also arise from generic, template-like videos that share little with any specific patient, so the privacy result is best read as 'no easy re-identification in this retrieval setup' rather than a guarantee against a trained re-identification adversary.
  • Because videos longer than 21 frames are subsampled in reverse chronological order, the generated sequence's temporal ordering need not follow the real fluorescein bolus phases; a direct check is whether the venous phase reliably precedes the late phase in generated clips.
  • If the model truly learns text-conditioned disease dynamics, the same architecture should work for other angiographic modalities, such as indocyanine green angiography, with minimal changes to the text encoder and prompt vocabulary.
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

5 major / 6 minor

Summary. This manuscript presents FFA Sora, a text-to-video diffusion transformer for generating 21-frame fundus fluorescein angiography (FFA) videos from textual clinical reports. The model combines a Wavelet-Flow Variational Autoencoder (WF-VAE) with a Diffusion Transformer (DiT) adapted from the Open-Sora Plan. The authors train on 3,625 FFA videos paired with 1,814 reports, randomly split 80/10/10 at the video level, and report FVD=329.78, LPIPS=0.48, VQAScore=0.61, BERTScore=0.35, average Recall@K=0.073, and a human rating of 1.57 (where 1 is best). The central claims are that the model accurately simulates disease features from text and preserves patient privacy, enabling synthetic FFA sharing for education and model training.

Significance. The application is novel and timely: text-to-video generation for FFA has clear educational and data-sharing value, and using a DiT for continuous FFA video generation is a reasonable extension of existing open-source video-diffusion methods. The authors should be credited for using established video-quality metrics, for making the base framework code available, and for attempting a domain-specific text-alignment evaluation. However, the reported evidence does not currently support the headline claims: the video-level random split with shared reports creates a leakage risk, the objective metrics lack baselines and uncertainty estimates, and the retrieval-based privacy evaluation is not a meaningful re-identification test. These issues are correctable with re-splitting and additional experiments, so I view this as a major revision rather than a rejection.

major comments (5)
  1. [Methods (Dataset) and Results (first paragraph)] The dataset is stated to contain 3,625 FFA videos paired with only 1,814 reports, and the split is described as "randomly selected" 80/10/10 with no patient-level or eye-level stratification. Because multiple videos from the same eye or patient share a single report, a random video-level split places videos from the same source into both training and test sets with high probability. This identity leakage can inflate FVD, LPIPS, VQAScore, and the human score through memorization of individual vascular patterns, and it invalidates the retrieval-based privacy claim because the test set is not disjoint from training identities. The authors should re-split at the patient or eye level and report the number of unique patients or eyes in each fold, or otherwise demonstrate that the report-to-video mapping does not cross the split.
  2. [Results, Table 1A] FVD=329.78, LPIPS=0.48, VQAScore=0.61, and BERTScore=0.35 are single point estimates with no confidence intervals, no multiple-seed variation, and no baseline values. The claim that these numbers indicate "excellent quality" and "satisfactory performance" requires comparison with, at minimum, the scores of real videos against themselves, a text-unconditioned or randomly conditioned generator, and a previously published video-generation model. Without such baselines, the absolute values are uninterpretable.
  3. [Objective Evaluation and Results (Figure 3)] The domain-specific alignment evaluation uses FFA-GPT (ref. 15) to translate generated videos back into text and then computes BERTScore against the input prompt. FFA-GPT is a model developed by the same group on the same dataset, so this is an internal text-video-text roundtrip rather than an external clinical benchmark. Moreover, BERTScore=0.35 is low in absolute terms, and no reference BERTScore is given for real videos under the same pipeline. The evaluation should be repeated with an independent report-generation model or with masked clinical readers, and reference scores should be reported.
  4. [Results, Table 1B (privacy evaluation)] The privacy claim rests on low Recall@K in an image-to-image retrieval task against ground-truth test videos. Low recall under a single feature extractor only shows that generated videos are not top-K nearest neighbors of the corresponding test videos in that embedding space; it does not test for training-set memorization, membership inference, or re-identification of patients whose videos appeared in training. Given the video-level split leakage described above, the reported Recall@K values do not establish "strong privacy-preserving performance." The authors should add a direct memorization test, such as nearest-neighbor search of generated videos against the full training set, and a membership-inference experiment.
  5. [Human Assessment] The human evaluation was performed by three authors (X.W., X.Y., Y.F.) on 50 generated videos, without blinding to the model outputs or to the study hypothesis, and without comparison against a baseline generator or against real videos. The resulting average score of 1.570 cannot support the claim of "satisfactory visual quality" as an unbiased clinical assessment. Independent masked raters, a pre-registered rating protocol, and inter-rater reliability statistics are needed.
minor comments (6)
  1. [Methods (frame interpolation)] The interpolation formula is written as F_interpolated = (1−t)×F_i − t×F_{i+1}, which contains a minus sign and would not produce a convex combination. This should be a plus sign: (1−t)×F_i + t×F_{i+1}.
  2. [Code availability] The Code availability section links only to the Open-Sora Plan repository, not to the FFA Sora implementation, preprocessing scripts, or model weights. Please provide the actual project code to make the results reproducible.
  3. [Results and Table 1B] The text reports an average Recall of 0.073, while Table 1B lists 0.07; please make these consistent.
  4. [Methods (Dataset)] The frame-preselection rule ("selected 21 frames in reverse chronological order") means the standardized videos emphasize venous and late phases; please clarify whether this temporal ordering matches the FFA examination protocol and how it affects the model's ability to generate the arterial phase.
  5. [Limitations] The second limitation in the Limitations section states that "our evaluation approach may not fully capture the broader generative quality." This is an important caveat that should be reflected in the abstract and Results claims, not only in the limitation paragraph.
  6. [Discussion] There is a typo "angioagraphy" in the Discussion; please proofread the manuscript.

Circularity Check

2 steps flagged · score 4.0 of 10

Domain-specific alignment is validated by a same-group, same-dataset FFA-GPT roundtrip, and the privacy conclusion restates the retrieval metric's definition; the core text-to-video training itself is not circular.

  1. self citation load bearing [Methods, Objective Evaluation; Results, Objective evaluation of generated FFA videos]
    "Based on previous development of FFA-GPT,15 a two-stage system which can generate FFA report automatically, our model was evaluated to determine whether the content of the generated videos aligns with the actual characteristics of retinal diseases as denoted by specified textual prompts, with Bidirectional Encoder Representations from Transformers Score (BERTScore) = 0.35 (Table 1A)."

    FFA-GPT is prior work by overlapping authors (Shi, He, Zhang are co-authors here) and was developed on the same FFA dataset. The alignment check is therefore a text-video-text roundtrip inside the authors' own pipeline: the original report is turned into video by FFA Sora, FFA-GPT turns that video back into text, and BERTScore compares the two texts. This does not validate the generated video against an independent clinical ground truth; it only checks consistency with an in-house report generator that shares data and authors with the model under test. The central claim that FFA Sora 'accurately simulates disease features from the input text' thus relies on a self-citation for its domain-specific evidence, and the score is presented as 'acceptable' without external benchmarks.

  2. self definitional [Results, Image retrieval demonstrated FFA Sora's excellent performance on privacy preserving; Table 1B caption]
    "The lower the Recall@K value, the smaller the overlapping part between the generated FFA videos and the ground-truth videos, demonstrating the better performance in preserving patient privacy. ... These relatively low Recall@K values indicate that FFA Sora effectively prevents the leakage of confidential image information during video generation."

    Privacy-preserving is defined in the Methods and Table caption as low Recall@K under a single image-retrieval model, and the Results then cite the same low Recall@K values as evidence that the model 'effectively prevents the leakage of confidential image information.' This is a definitional loop: the metric is operationalized as 'lower retrieval overlap equals more privacy,' so the conclusion restates the operationalization rather than testing re-identification risk. No formal privacy adversary, membership-inference test, or patient-level analysis is provided, and the 'content-preserving' half of the goal is not measured in the same retrieval design. The claimed privacy guarantee therefore has no content beyond the chosen retrieval metric.

full rationale

The core training and generation pipeline is not circular: FFA Sora learns a conditional distribution over videos given text from paired FFA-report data, and the reported FVD, LPIPS, and VQAScore are standard external metrics applied to held-out test videos. However, the domain-specific alignment validation is self-referential: it passes the generated video through FFA-GPT, a report generator published by the same group and trained on the same FFA dataset, and then compares the round-tripped text with the original prompt. That does not independently establish that the generated video contains the described disease features; it only establishes consistency with the authors' own report-generation model. Similarly, the privacy claim is an operational tautology: the paper equates 'privacy-preserving' with low Recall@K and then cites the low Recall@K values as proof of privacy, without testing against a realistic adversary or a formal re-identification threat model. A separate concern, not scored as circularity, is the absence of a patient- or eye-level split: given 3,625 videos paired with only 1,814 reports, a random video-level split may place videos sharing a report (and possibly the same eye) in both training and testing, which is a validity risk for the reported test metrics. Taken together, the central model has independent content, but the two key validation claims—text alignment and privacy preservation—are substantially supported by self-referential or definitional evidence, so the paper merits a partial circularity score of 4.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several unproven domain assumptions: report-video pairing specificity, FFA-GPT as a faithful evaluator, low retrieval as privacy, and unbiased author raters. No new physical or mathematical entities are introduced. Model hyperparameters are unreported, and the two hand-chosen preprocessing constants (vessel-area cutoff and 21-frame length) are free choices that shape the data distribution.

free parameters (3)
  • vessel_area_ratio_filter_threshold = 0.005
    Videos with frame vessel area ratio below 0.005 were excluded; the cutoff is chosen by hand, not justified, and it shapes the training distribution.
  • standardized_frame_count = 21
    All videos were truncated or linearly interpolated to exactly 21 frames; this arbitrary choice removes early arterial-phase dynamics for longer videos and may favor late-phase appearance.
  • train_validation_test_split = 80/10/10
    The dataset split is chosen by the authors; no stratification by disease or patient is described, and with 1,814 reports paired to 3,625 videos, patient-level leakage across splits is possible.
assumptions (4)
  • domain assumption Each FFA report text is an accurate and sufficiently specific description of the paired FFA video(s).
    Training relies on text-video pairs from retrospective clinical records; the paper reports 3,625 videos paired with 1,814 reports, so the mapping is many-to-one and prompt specificity is not verified.
  • domain assumption FFA-GPT yields faithful text descriptions of generated videos for evaluation.
    FFA-GPT (ref 15) is a same-group model trained on the same data; using it as a ground-truth reporter assumes its errors do not bias the BERTScore comparison.
  • ad hoc to paper Low image-retrieval Recall@K is a valid proxy for privacy preservation.
    The privacy conclusion is inferred from low Recall@K (0.073), but no formal privacy definition, membership-inference test, or re-identification attack is performed.
  • ad hoc to paper Author-ophthalmologists provide unbiased subjective quality ratings.
    Three of the raters are co-authors (X.W., X.Y., Y.F.), and no external clinician validation or inter-rater reliability metric is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FFA Sora, video generation as fundus fluorescein angiography simulator." pith.science (2026). https://pith.science/paper/GET33JBD

@misc{pith2026241217346,
  author       = {Pith},
  title        = {Pith review of: FFA Sora, video generation as fundus fluorescein angiography simulator},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GET33JBD}},
  note         = {Machine review of arXiv:2412.17346}
}
read the original abstract

Fundus fluorescein angiography (FFA) is critical for diagnosing retinal vascular diseases, but beginners often struggle with image interpretation. This study develops FFA Sora, a text-to-video model that converts FFA reports into dynamic videos via a Wavelet-Flow Variational Autoencoder (WF-VAE) and a diffusion transformer (DiT). Trained on an anonymized dataset, FFA Sora accurately simulates disease features from the input text, as confirmed by objective metrics: Frechet Video Distance (FVD) = 329.78, Learned Perceptual Image Patch Similarity (LPIPS) = 0.48, and Visual-question-answering Score (VQAScore) = 0.61. Specific evaluations showed acceptable alignment between the generated videos and textual prompts, with BERTScore of 0.35. Additionally, the model demonstrated strong privacy-preserving performance in retrieval evaluations, achieving an average Recall@K of 0.073. Human assessments indicated satisfactory visual quality, with an average score of 1.570(scale: 1 = best, 5 = worst). This model addresses privacy concerns associated with sharing large-scale FFA data and enhances medical education.

Figures

Figures reproduced from arXiv: 2412.17346 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p016_1.png] view at source ↗
Figure 2
Figure 2. [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 1
Figure 1. Development of the text-to-video diffusion model. This figure illustrates a text-to￾video framework for generating Fluorescein Fundus Angiography sequences using a Denoising Diffusion Implicit Transformer. Medical text prompts are encoded by the Text Encoder into a semantic embedding to guide video generation. Input FFA sequences are encoded into latent representations by the Encoder and progressively refined throug… view at source ↗
Figures from the paper (1 more)
Figure 2
Figure 2. Figure 2: Representative examples of generated FFA videos. BRVO = branch retinal vein occlusion, CRVO = central retinal vein occlusion, NPDR = non-proliferative diabetic retinopathy, PCV = polypoidal choroidal vasculopathy, AMD = age-related macular degeneration, CSC = central s…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 34 canonical work pages

  1. [1]

    Burns, S.A., Elsner, A.E., and Gast, T.J. (2021). Imaging the Retinal Vasculature. Annual Review of Vision Science 7, 129-153. 10.1146/annurev-vision-093019-113719

    work page doi:10.1146/annurev-vision-093019-113719 2021

  2. [2]

    Sulzbacher, F., Pollreisz, A., Kaider, A., Kickinger, S., Sacu, S., and Schmidt-Erfurth, U. (2017). Identification and clinical role of choroidal neovascularization characteristics based on optical coherence tomography angiography. Acta Ophthalmol 95, 414-420. 10.1111/aos.13364

    work page doi:10.1111/aos.13364 2017

  3. [3]

    Do, D.V. (2013). Detection of new-onset choroidal neovascularization. Curr Opin Ophthalmol 24, 244-247. 10.1097/ICU.0b013e32835fd7dd

    work page doi:10.1097/icu.0b013e32835fd7dd 2013

  4. [4]

    Coscas, G.J., Lupidi, M., Coscas, F., Cagini, C., and Souied, E.H. (2015). OPTICAL COHERENCE TOMOGRAPHY ANGIOGRAPHY VERSUS TRADITIONAL MULTIMODAL IMAGING IN ASSESSING THE ACTIVITY OF EXUDATIVE AGE-RELATED MACULAR DEGENERATION: A New Diagnostic Challenge. Retina 35, 2219-2228. 10.1097/iae.0000000000000766

    work page doi:10.1097/iae.0000000000000766 2015

  5. [5]

    Sabanayagam, C., Banu, R., Chee, M.L., Lee, R., Wang, Y.X., Tan, G., Jonas, J.B., Lamoureux, E.L., Cheng, C.Y., Klein, B.E.K., et al. (2019). Incidence and progression of diabetic retinopathy: a systematic review. Lancet Diabetes Endocrinol 7, 140-149. 10.1016/s2213-8587(18)30128-1

    work page doi:10.1016/s2213-8587(18)30128-1 2019

  6. [6]

    Tabuchi, H., Engelmann, J., Maeda, F., Nishikawa, R., Nagasawa, T., Yamauchi, T., Tanabe, M., Akada, M., Kihara, K., Nakae, Y., et al. (2024). Using artificial intelligence to improve human performance: efficient retinal disease detection training with synthetic images. The British Journal of Ophthalmology 108, 1430-1435. 10.1136/bjo-2023-324923

    work page doi:10.1136/bjo-2023-324923 2024

  7. [7]

    Correspondence: Dr

    Lead contact. Correspondence: Dr. Danli Shi, The Hong Kong Polytechnic University, Hong Kong, China. Email: danli.shi@polyu.edu.hk (Lead contact) Prof. Mingguang He, Chair Professor of Experimental Ophthalmology, The Hong Kong Polytechnic University, Hong Kong, China. Email: mingguang.he@polyu.edu.hk Summary 2 Fundus fluorescein angiography (FFA) is criti...

    work page 2016

  8. [8]

    Price, W.N., 2nd, and Cohen, I.G. (2019). Privacy in the age of medical big data. Nat Med 25, 37-43. 10.1038/s41591-018-0272-7

    doi:10.1038/s41591-018-0272-7 2019

Show all 46 references
  1. [9]

    O'Callaghan, J. (2024). How OpenAI’s text-to-video tool Sora could change science – and society. Nature 627, 475-476. 10.1038/d41586-024-00661-0

    2024 doi

  2. [10]

    Dhariwal, P., and Nichol, A. (2021). Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794

    2021

  3. [11]

    Bluethgen, C., Chambon, P., Delbrouck, J.-B., van der Sluijs, R., Połacin, M., Zambrano Chaves, J.M., Abraham, T.M., Purohit, S., Langlotz, C.P., and Chaudhari, A.S. (2024). A vision-language foundation model for the generation of realistic chest X-ray images. Nature Biomedica...

    2024 doi

  4. [12]

    Dorjsembe, Z., Pao, H.-K., Odonchimed, S., and Xiao, F. (2024). Conditional Diffusion Models for Semantic 3D Brain MRI Synthesis. IEEE Journal of Biomedical and Health Informatics 28, 4084-4093. 10.1109/JBHI.2024.3385504

    2024

  5. [13]

    Chen, R., Zhang, W., Liu, B., Chen, X., Xu, P., Liu, S., He, M., and Shi, D. (2024). EyeDiff: text-to-image diffusion model improves rare eye disease diagnosis. arXiv preprint arXiv:2411.10004

    2024 arXiv

  6. [14]

    Li, C., Liu, H., Liu, Y., Feng, B.Y., Li, W., Liu, X., Chen, Z., Shao, J., and Yuan, Y. (2024). Endora: Video Generation Models as Endoscopy Simulators

    2024

  7. [15]

    Sener, F., Saraf, R., and Yao, A. (2023). Transferring Knowledge From Text to Video: Zero-Shot Anticipation for Procedural Actions. IEEE transactions on pattern analysis and 19 machine intelligence 45, 7836-7852. 10.1109/TPAMI.2022.3218596

    2023

  8. [16]

    Chen, X., Zhang, W., Xu, P., Zhao, Z., Zheng, Y., Shi, D., and He, M. (2024). FFA-GPT: an automated pipeline for fundus fluorescein angiography interpretation and question-answer. npj Digital Medicine 7, 1-9. 10.1038/s41746-024-01101-z

    2024 doi

  9. [17]

    Lin, B., Ge, Y., Cheng, X., Li, Z., Zhu, B., Wang, S., He, X., Ye, Y., Yuan, S., Chen, L., et al. (2024). Open-Sora Plan: Open-Source Large Video Generation Model

    2024

  10. [18]

    Li, Z., Lin, B., Ye, Y., Chen, L., Cheng, X., Yuan, S., and Yuan, L. (2024). WF-VAE: Enhancing Video VAE by Wavelet-Driven Energy Flow for Latent Video Diffusion Model

    2024

  11. [19]

    Peebles, W.S., and Xie, S. (2022). Scalable Diffusion Models with Transformers. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 4172-4182

    2022

  12. [20]

    Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., and Gelly, S. (2019). FVD: A new metric for video generation

    2019

  13. [21]

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., and Wang, O. (2018). The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society

    2018

  14. [22]

    Lin, Z., Pathak, D., Li, B., Li, J., Xia, X., Neubig, G., Zhang, P., and Ramanan, D. (2024). Evaluating Text-to-Visual Generation with Image-to-Text Generation

    2024

  15. [23]

    Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., and Artzi, Y. (2019). BERTScore: Evaluating Text Generation with BERT. ArXiv abs/1904.09675

    2019 arXiv

  16. [24]

    Zhang, W., Yang, J., Chen, R., Huang, S., Xu, P., Chen, X., Lu, S., Cao, H., He, M., and Shi, D. (2024). Fundus to Fluorescein Angiography Video Generation as a Retinal Generative Foundation Model. ArXiv abs/2410.13242

    2024 arXiv

  17. [25]

    Zhang, W., Huang, S., Yang, J., Chen, R., Ge, Z., Zheng, Y., Shi, D., and He, M. (2024). Fundus2Video: Cross-Modal Angiography Video Generation from Static Fundus Photography with Clinical Knowledge Guidance. held in Cham, (Springer Nature Switzerland), pp. 689-699

    2024

  18. [26]

    Tian, Y., Fan, L., Isola, P., Chang, H., and Krishnan, D. (2023). StableRep: Synthetic Images from Text-to-Image Models Make Strong Visual Representation Learners. ArXiv abs/2306.00984

    2023 arXiv

  19. [27]

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. (2021). High-Resolution Image Synthesis with Latent Diffusion Models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 10674-10685

    2021

  20. [28]

    Croitoru, F.A., Hondru, V., Ionescu, R.T., and Shah, M. (2023). Diffusion Models in Vision: A Survey. IEEE Trans Pattern Anal Mach Intell 45, 10850-10869. 10.1109/tpami.2023.3261988

    2023

  21. [29]

    He, S., Joseph, S., Bulloch, G., Jiang, F., Kasturibai, H., Kim, R., Ravilla, T.D., Wang, Y., Shi, D., and He, M. (2023). Bridging the Camera Domain Gap With Image-to-Image Translation Improves Glaucoma Diagnosis. Transl Vis Sci Technol 12, 20-20. 10.1167/tvst.12.12.20

    2023 doi

  22. [30]

    Song, F., Zhang, W., Zheng, Y., Shi, D., and He, M. (2023). A deep learning model for generating fundus autofluorescence images from color fundus photography. Adv Ophthalmol Pract Res 3, 192-198. 10.1016/j.aopr.2023.11.001

    2023 doi

  23. [31]

    Shi, D., He, S., Yang, J., Zheng, Y., and He, M. (2024). One-shot Retinal Artery and Vein Segmentation via Cross-modality Pretraining. Ophthalmol Sci 4, 100363. 20 10.1016/j.xops.2023.100363

    2024

  24. [32]

    Chen, R., Xu, K., Zheng, K., Zhang, W., Lu, Y., Shi, D., and He, M. (2024). Generating Multi-frame Ultrawide-field Fluorescein Angiography from Ultrawide-field Color Imaging Improves Diabetic Retinopathy Stratification. arXiv preprint arXiv:2408.10636

    2024

  25. [33]

    Shi, D., Zhang, W., He, S., Chen, Y., Song, F., Liu, S., Wang, R., Zheng, Y., and He, M. (2023). Translation of Color Fundus Photography into Fluorescein Angiography Using Deep Learning for Enhanced Diabetic Retinopathy Screening. Ophthalmol Sci 3, 100401. 10.1016/j.xops.2023.100401

    2023

  26. [34]

    Zhang, W., Huang, S., Yang, J., Chen, R., Ge, Z., Zheng, Y., Shi, D., and He, M. (2024). Fundus2Video: Cross-Modal Angiography Video Generation from Static Fundus Photography with Clinical Knowledge Guidance. held in Morocco, (Springer Nature Switzerland), pp. 689-699

    2024

  27. [35]

    Chen, R., Zhang, W., Song, F., Yu, H., Cao, D., Zheng, Y., He, M., and Shi, D. (2024). Translating color fundus photography to indocyanine green angiography using deep-learning for age-related macular degeneration screening. npj Digital Medicine 7,

    2024

  28. [36]

    10.1038/s41746-024-01018-7

    doi

  29. [37]

    Guo, X., Wen, H., Hao, H., Zhao, Y., Meng, Y., Liu, J., Zheng, Y., Chen, W., and Zhao, Y. (2024). Randomness-restricted Diffusion Model for Ocular Surface Structure Segmentation. IEEE Trans Med Imaging Pp. 10.1109/tmi.2024.3494762

    2024

  30. [38]

    Aktas, B., Ates, D.D., Duzyel, O., and Gumus, A. (2024). Diffusion-based data augmentation methodology for improved performance in ocular disease diagnosis using retinography images. International Journal of Machine Learning and Cybernetics. 10.1007/s13042-024-02485-w

    2024 doi

  31. [39]

    Peebles, W., and Xie, S. (2023). Scalable diffusion models with transformers. pp. 4195-4205

    2023

  32. [40]

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. (2020). An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. ArXiv abs/2010.11929

    2020 arXiv

  33. [41]

    Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., and Zhu, J. (2022). All are Worth Words: A ViT Backbone for Diffusion Models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 22669-22679

    2022

  34. [42]

    Yang, Y., Chen, X., and Lin, H. (2024). Privacy preserving technology in ophthalmology. Current Opinion in Ophthalmology 35, 431-437. 10.1097/ICU.0000000000001087

    2024 doi

  35. [43]

    Aa, S., M, G., C, A.-R., Gb, L.E., A, M.-G., Mj, E.-C., and Cl, P.-C. (2024). Privacy-preserving federated machine learning on FAIR health data: A real-world application. Computational and structural biotechnology journal

    2024

  36. [44]

    10.1016/j.csbj.2024.02.014

    2024 doi

  37. [45]

    Collado-Mesa, F., Alvarez, E., and Arheart, K. (2018). The Role of Artificial Intelligence in Diagnostic Radiology: A Survey at a Single Radiology Residency Training Program. J Am Coll Radiol 15, 1753-1757. 10.1016/j.jacr.2017.12.021

    2018 doi

  38. [46]

    Coyner, A.S., Chen, J.S., Chang, K., Singh, P., Ostmo, S., Chan, R.V.P., Chiang, M.F., Kalpathy-Cramer, J., Campbell, J.P., and Consortium, I.a.I.i.R.o.P. (2022). Synthetic Medical Images for Robust, Privacy-Preserving Training of Artificial Intelligence: Application to Retino...

    2022

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.