Pith. sign in

REVIEW 2 major objections 6 minor 71 references

The paper introduces OCT-Bench and argues that current multimodal models cannot yet be trusted for OCT-based clinical reasoning.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 20:25 UTC pith:LZMZJUOB

load-bearing objection A genuinely new and useful OCT benchmark with a plausible capability hierarchy, but the unquantified expert QC and missing release details mean the headline numbers should be read as provisional. the 2 major comments →

arxiv 2607.16609 v1 pith:LZMZJUOB submitted 2026-07-18 cs.CV cs.CL

Can Multimodal Large Language Models Understand OCT?

Baochen Fu , Wenzhi Deng , Baihao Jin , Yang Li , Zihan Nie , Kailin Jiang , Yuntao Du , Weiye Song This is my paper
classification cs.CV cs.CL
keywords OCT-Benchmultimodal large language modelsoptical coherence tomographymedical image understandingbenchmarkclinical reasoningretinal imagingvision-language models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that current multimodal large language models cannot yet carry out the complete chain of OCT image interpretation, from seeing retinal details to reasoning about diagnosis and treatment. To test that, it builds OCT-Bench, a benchmark of 10,076 expert-checked multiple-choice questions on 4,137 OCT images, organized into Perception, Cognition, and Reasoning levels that mirror the clinical workflow. The authors report that the best model reaches only 62.0% overall accuracy, with a steep drop from a 75.8% perception ceiling to a 42.9% reasoning ceiling. They also find that neither larger model scale nor medical-domain training consistently closes the gap. If true, the benchmark gives the community a fine-grained way to locate whether an AI failure is visual, cognitive, or clinical-reasoning in origin.

Core claim

On OCT-Bench, no model comes close to reliable OCT interpretation: the best overall accuracy is 62.0%, and the highest Reasoning score is only 42.9%. The paper's central claim is that performance degrades systematically as the required capability advances from visual perception to medical cognition to clinical reasoning, and that neither larger scale nor medical-domain training resolves this. Beneath the aggregate numbers, models are near-saturated on trivial perception tasks but collapse on fine-grained morphology, retinal layer identification, and most disease-diagnosis tasks, especially diseases requiring subtle differential reasoning such as AMD, glaucoma, and retinal vascular occlusion.

What carries the argument

The load-bearing object is OCT-Bench's hierarchical capability taxonomy, which decomposes OCT understanding into three dimensions—Perception, Cognition, and Reasoning—further divided into nine capability groups and 20 fine-grained tasks such as reflectivity analysis, layer identification, spatial relations, stage classification, and follow-up adjustment. The taxonomy carries the argument: by separating levels, the authors can attribute a wrong answer to a particular stage and show how errors propagate from low-level perception to high-level reasoning. The construction pipeline, including a two-stage expert quality-control step, is what makes the 10,076 questions a credible measuring instrume

Load-bearing premise

OCT-Bench's accuracy scores are only meaningful as measures of OCT understanding if the expert quality-control step truly removed every question that could be answered from language alone, option patterns, or common clinical priors; the paper asserts this cleanup but reports no inter-rater statistics or remaining-cue analysis.

What would settle it

Re-run the same 20 models on OCT-Bench with option letters shuffled per question; if Reasoning scores rise sharply or vary with option order, the low reasoning numbers are partly a test-taking artifact. A second check is to have two independent expert panels re-flag any question whose answer is guessable without the image; if substantial numbers survive, the perception-to-reasoning gap is inflated.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Current multimodal models, as measured by OCT-Bench, are not reliable for clinical OCT interpretation: the best model answers roughly 38 percent of questions incorrectly.
  • Overall accuracy masks a consistent capability bottleneck: performance drops from perception to cognition to reasoning, so a single score cannot tell where a model fails.
  • Scaling model size and medical-domain training do not consistently improve all three levels; in some families, scaling raises cognition scores while leaving reasoning almost unchanged.
  • Fine-grained visual description and layer identification are the weakest points, with the best models scoring below roughly 57 percent, indicating a visual grounding gap rather than only a knowledge gap.
  • Disease diagnosis is uneven: common patterns such as macular hole and central serous chorioretinopathy are recognized far better than glaucoma or retinal vascular occlusion, suggesting category-dependent shortcuts.

Where Pith is reading between the lines

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

  • [Editorial inference] If the 42.9 percent reasoning ceiling holds up under stricter cue-leakage controls, OCT-Bench becomes a safety gate: no multimodal model should be used for autonomous OCT-based triage or treatment advice until it clears a much higher reasoning bar.
  • [Editorial inference] The hierarchy suggests a training remedy: because reasoning failures may propagate from earlier perceptual and cognitive errors, targeted supervision on mid-level tasks such as layer identification and lesion classification could yield more reasoning gain than simply scaling parameters.
  • [Editorial inference] A testable extension is to shuffle option order and re-run the same models; if Reasoning scores are stable, the low ceiling reflects genuine inference limits rather than option-position cues.
  • [Editorial inference] The disease-category imbalance implies that aggregate diagnosis accuracy overstates clinical safety; deployment evaluations should report per-disease accuracy, especially for low-prevalence conditions like retinal artery or vein occlusion.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The manuscript introduces OCT-Bench, a benchmark of 10,076 multiple-choice questions constructed from 4,137 OCT images across seven public datasets, organized into a hierarchical taxonomy (Perception, Cognition, Reasoning; 9 capability groups; 20 fine-grained tasks). The authors evaluate 20 MLLMs, including proprietary, open-source general-purpose, and medical-domain models, under a zero-shot MCQ protocol. The reported results show a best overall accuracy of 62.0% (GPT-5.4-mini), with the highest Perception score 75.8%, Cognition 64.2%, and Reasoning 42.9%, and the paper concludes that current MLLMs remain substantially short of reliable OCT understanding. It further reports that neither medical-domain adaptation nor increased model scale yields consistent improvements across capability levels. The main contributions are the benchmark, the clinically motivated taxonomy, and the systematic multi-model evaluation with fine-grained per-task results.

Significance. If the benchmark's validity is confirmed, OCT-Bench is a timely and useful contribution: it is substantially larger and more fine-grained than existing OCT-relevant benchmarks, the code and data are promised to be public, and the evaluation covers a broad set of 20 models. The perception-to-reasoning gradient is consistent across model families, and the per-task breakdown provides actionable information about where models fail. The main risk is construct validity: the expert quality-control step is asserted but not quantified, and several Perception tasks measure annotation metadata rather than intrinsic OCT content. These issues are addressable and do not, at this stage, require rejection, but they must be fixed before the benchmark can support strong conclusions about clinical OCT understanding.

major comments (2)
  1. [Benchmark Construction, Step 5; Data Analysis] The paper asserts that expert quality control revises/removes samples with insufficient visual evidence, ambiguity, or language-only cues, but no quantitative support is reported: number of experts, qualification criteria, inter-rater agreement (e.g., Cohen's kappa), or counts of generated/rejected/revised questions. This is load-bearing because the central claim that the best model is 'far from reliable OCT understanding' and that Reasoning is a bottleneck assumes the benchmark cannot be solved by textual priors or option bias. Please report these QC statistics and run post-hoc analyses: (i) distribution of correct options across A-D, (ii) a text-only baseline that is given the question and options but not the image, and (iii) per-option or item-level checks for lexical cues. Without these, the reported 42.9% Reasoning ceiling could be an artifact of question construction.
  2. [OCT-Bench; Figures 1-2, Table 3 (T02, T07, T08)] Several 'Perception' tasks measure metadata of the annotation overlay rather than intrinsic OCT content. T02 asks the color of the annotation box; T07 asks which annotation box is larger; T08 asks the relative position of two annotation boxes. These are near-saturated (e.g., T01/T02 ~98% for most models) and inflate the Perception dimension and therefore the headline perception-to-reasoning gradient. Please either remove/separate such annotation-metadata tasks from the clinical Perception score or demonstrate that the gradient persists when Perception is computed only from tasks relying on OCT features (T03-T05, T09-T10, etc.). At minimum, report a 'Perception-internal' score excluding T01/T02/T07/T08.
minor comments (6)
  1. [Figure 1] There is a typo: 'Congnition' should be 'Cognition'. Also, the L1/L2/L3 notation in Figure 2 is not explicitly defined in the caption; clarify the level hierarchy for readers.
  2. [Table 1] The counts in the rightmost columns are hard to parse (e.g., the LMOD row). State explicitly what each number represents, e.g., number of tasks per capability dimension.
  3. [Fine-grained Analysis, Figures 4-5] Report the number of questions per disease category (Figure 5) and per anatomical region (Figure 4). Some accuracies, such as RVO at 2.5%, may be based on very few items, and without sample sizes the category-level conclusions are difficult to evaluate.
  4. [Main Results; Table 2] No confidence intervals or significance tests are reported for the headline gaps (e.g., the 32.9-point perception-to-reasoning drop). Add at least binomial confidence intervals or bootstrap intervals so readers can judge whether the gaps are statistically reliable.
  5. [Evaluation Strategy] The protocol counts invalid responses as incorrect. Please report per-model invalid-response rates, since models that fail to follow the 'output only the option letter' instruction are disproportionately penalized and this may affect model rankings.
  6. [Data Analysis] The text says that 'detailed statistics and distribution analyses are provided in the appendix,' but the submitted version contains no appendix. Include those statistics or remove the reference.

Circularity Check

0 steps flagged

No circularity: OCT-Bench results are independent empirical measurements, not quantities defined by fitted parameters or by the benchmark's own construction.

full rationale

The paper's central contribution is a benchmark and an evaluation. The reported accuracies (e.g., 62.0% overall, 42.9% best Reasoning) are measurements of model outputs against expert-verified ground-truth answers; they are not defined in terms of any fitted parameter, and no equation in the paper reduces a prediction to an input. The taxonomy (Perception/Cognition/Reasoning) is an organizing scheme, not a derivation. The use of GPT-4o to draft questions is a construction choice, but the paper's Step 5 expert quality control is an external check, and even if one doubts the sufficiency of that check, that is a validity concern rather than circularity: the model scores are not statistically forced by the question-generation process. Self-citations in related work (e.g., Fu et al. 2026, Jiang et al. 2026a) support contextual claims about other benchmarks and are not load-bearing for the measured performance results. No self-definitional step, fitted-input-called-prediction step, or author-imported uniqueness theorem appears. The weakest point—unquantified expert QC—affects the benchmark's evidentiary value, not the circularity of the derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

The central claim is a benchmark, so the ledger is dominated by assumptions about data quality, question validity, and evaluation fairness rather than fitted parameters. There are no fitted constants or invented entities. The key uncertainty is whether the expert-QC and task-isolation assumptions hold.

axioms (4)
  • domain assumption The seven public OCT datasets provide sufficient, unbiased image/annotation coverage to construct a representative benchmark.
    The benchmark's validity depends on the source datasets (OCT5k, OIMHS, OCT-C8, AMD-SD, OCTDL, MMC-AMD, GOALS) being representative and their annotations correct; the paper standardizes them but does not verify external validity.
  • domain assumption The expert quality-control step removes all ambiguous, unanswerable, and language-only-cue questions.
    Stated in Step 5 (Expert Quality Control) but not quantified; if false, accuracy scores could reflect shortcuts rather than OCT understanding.
  • domain assumption Each task in the taxonomy isolates the intended capability without requiring other capabilities.
    The paper says tasks are designed to 'minimize overlap among different capabilities,' but no task-purity analysis (e.g., correlation or ablation) is provided.
  • domain assumption The 20 evaluated models are given prompts that fairly elicit their abilities under a unified zero-shot setting.
    The exact prompt template and inference hyperparameters are not reported, so it is assumed the evaluation is fair across models.

pith-pipeline@v1.3.0-alltime-deepseek · 15870 in / 10476 out tokens · 106822 ms · 2026-08-01T20:25:30.192622+00:00 · methodology

0 comments
read the original abstract

Optical coherence tomography (OCT) imaging is essential for the diagnosis and treatment of retinal diseases. Although multimodal large language models (MLLMs) have demonstrated considerable potential in medical image analysis, existing benchmarks largely reduce OCT understanding to coarse-grained disease classification or isolated visual question answering, leaving the complete cognitive process from visual perception to clinical reasoning insufficiently evaluated. To address this limitation, we introduce OCT-Bench, a comprehensive benchmark dedicated to OCT image understanding. OCT-Bench comprises 10,076 high-quality multiple-choice questions constructed from 4,137 OCT images across seven public datasets. Following the real-world clinical interpretation workflow, we establish a hierarchical capability taxonomy consisting of 20 fine-grained tasks across three dimensions: Perception, Cognition, and Reasoning. These tasks cover a broad range of capabilities, including imaging attributes, retinal anatomy, lesion characteristics, spatial relationships, disease assessment, therapeutic decision-making, and prognostic management. We systematically evaluate 20 representative MLLMs, including proprietary models, open-source general-purpose models, and medical-domain models. Experimental results demonstrate that current models remain substantially short of reliable OCT understanding. Moreover, neither medical-domain adaptation nor increased model scale consistently improves performance across capability levels. OCT-Bench enables comprehensive and fine-grained evaluation of MLLMs, providing a foundation for identifying capability bottlenecks and advancing clinically grounded OCT understanding.

Figures

Figures reproduced from arXiv: 2607.16609 by Baihao Jin, Baochen Fu, Kailin Jiang, Weiye Song, Wenzhi Deng, Yang Li, Yuntao Du, Zihan Nie.

Figure 1
Figure 1. Figure 1: Overview of OCT-Bench. tation (Chen et al. 2026). OCT analysis does not proceed directly from an image to a diagnosis; rather, it follows a progressive pathway from visual perception through medi￾cal cognition to clinical reasoning. Existing evaluations of￾ten conflate these levels within a single task. When a model makes an incorrect prediction, it is therefore difficult to de￾termine whether the failure … view at source ↗
Figure 2
Figure 2. Figure 2: Task taxonomy of OCT-Bench. imaging modality for retinal disease diagnosis, OCT requires models to recognize fine-grained retinal structures and le￾sions while integrating medical knowledge for clinical rea￾soning. To fill this gap, OCT-Bench systematically evaluates MLLMs on OCT image understanding across three progres￾sive levels: visual perception, medical cognition, and clinical reasoning. OCT-Bench Hi… view at source ↗
Figure 3
Figure 3. Figure 3: Construction pipeline of OCT-Bench. The pipeline comprises data collection, evaluation task design, medical knowl [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Disease-category-level diagnosis accuracy for task [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Case study of Region Identification. 20 fine-grained tasks, enabling comprehensive evaluation of overall performance and capability bottlenecks. Evaluations of 20 representative MLLMs reveal that re￾liable OCT understanding remains challenging. The best model achieves only 62.0% overall accuracy, with perfor￾mance dropping markedly from perception to reasoning. Models consistently struggle with fine-graine… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

71 extracted references · 18 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2410.21276 , year=

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:2507.06261 , year=

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    Pith/arXiv arXiv

  3. [3]

    2025 , month = aug, note =

    Grok 4 Model Card , author =. 2025 , month = aug, note =

    2025

  4. [4]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Improved baselines with visual instruction tuning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  5. [5]

    arXiv preprint arXiv:2407.07895 , year=

    LLaVA-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models , author=. arXiv preprint arXiv:2407.07895 , year=

    Pith/arXiv arXiv

  6. [6]

    Microsoft GenAI , year=

    New models added to the Phi-3 family, available on Microsoft Azure , author=. Microsoft GenAI , year=

  7. [7]

    arXiv preprint arXiv:2412.08905 , year=

    Phi-4 technical report , author=. arXiv preprint arXiv:2412.08905 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2412.05271 , year=

    Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling , author=. arXiv preprint arXiv:2412.05271 , year=

    Pith/arXiv arXiv

  9. [9]

    5-vl technical report , author=

    Qwen2. 5-vl technical report , author=. arXiv preprint arXiv:2502.13923 , year=

    Pith/arXiv arXiv

  10. [10]

    arXiv preprint arXiv:2403.08295 , year=

    Gemma: Open models based on gemini research and technology , author=. arXiv preprint arXiv:2403.08295 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2502.09838 , year=

    Healthgpt: A medical large vision-language model for unifying comprehension and generation via heterogeneous knowledge adaptation , author=. arXiv preprint arXiv:2502.09838 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2507.05201 , year=

    Medgemma technical report , author=. arXiv preprint arXiv:2507.05201 , year=

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2506.07044 , year=

    Lingshu: A generalist foundation model for unified multimodal medical understanding and reasoning , author=. arXiv preprint arXiv:2506.07044 , year=

    Pith/arXiv arXiv

  14. [14]

    arXiv preprint arXiv:2510.08668 , year=

    Hulu-med: A transparent generalist model towards holistic medical vision-language understanding , author=. arXiv preprint arXiv:2510.08668 , year=

    arXiv

  15. [15]

    arXiv preprint arXiv:2602.23363 , year=

    Medix-r1: Open ended medical reinforcement learning , author=. arXiv preprint arXiv:2602.23363 , year=

    arXiv

  16. [16]

    arXiv preprint arXiv:2511.00916 , year=

    Fleming-VL: Towards Universal Medical Visual Reasoning with Multimodal LLMs , author=. arXiv preprint arXiv:2511.00916 , year=

    arXiv

  17. [17]

    European conference on computer vision , pages=

    Mmbench: Is your multi-modal model an all-around player? , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  18. [18]

    International Conference on Learning Representations , volume=

    Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans? , author=. International Conference on Learning Representations , volume=

  19. [19]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Unk-vqa: A dataset and a probe into the abstention ability of multi-modal large models , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2024 , publisher=

    2024

  20. [20]

    arXiv preprint arXiv:2401.11943 , year=

    Benchmarking large multimodal models against common corruptions , author=. arXiv preprint arXiv:2401.11943 , year=

    Pith/arXiv arXiv

  21. [21]

    International Conference on Learning Representations , volume=

    Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts , author=. International Conference on Learning Representations , volume=

  22. [22]

    arXiv preprint arXiv:2307.16125 , year=

    Seed-bench: Benchmarking multimodal llms with generative comprehension , author=. arXiv preprint arXiv:2307.16125 , year=

    Pith/arXiv arXiv

  23. [23]

    Ophthalmology science , volume=

    Evaluating the performance of ChatGPT in ophthalmology: an analysis of its successes and shortcomings , author=. Ophthalmology science , volume=. 2023 , publisher=

    2023

  24. [24]

    EBioMedicine , volume=

    Benchmarking large language models’ performances for myopia care: a comparative analysis of ChatGPT-3.5, ChatGPT-4.0, and Google Bard , author=. EBioMedicine , volume=. 2023 , publisher=

    2023

  25. [25]

    arXiv preprint arXiv:2502.14260 , year=

    Eyebench: A call for more rigorous evaluation of retinal image enhancement , author=. arXiv preprint arXiv:2502.14260 , year=

    Pith/arXiv arXiv

  26. [26]

    European Conference on Computer Vision , pages=

    Ophnet: A large-scale video benchmark for ophthalmic surgical workflow understanding , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  27. [27]

    Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    Lmod: A large multimodal ophthalmology dataset and benchmark for large vision-language models , author=. Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    2025

  28. [28]

    arXiv preprint arXiv:2603.15117 , year=

    Mmku-bench: A multimodal update benchmark for diverse visual knowledge , author=. arXiv preprint arXiv:2603.15117 , year=

    arXiv

  29. [29]

    science , volume=

    Optical coherence tomography , author=. science , volume=. 1991 , publisher=

    1991

  30. [30]

    Advances in Neural Information Processing Systems , volume=

    Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks , author=. Advances in Neural Information Processing Systems , volume=

  31. [31]

    Findings of the association for computational linguistics: ACL 2024 , pages=

    The revolution of multimodal large language models: A survey , author=. Findings of the association for computational linguistics: ACL 2024 , pages=

    2024

  32. [32]

    IEEE Transactions on Circuits and Systems for Video Technology , year=

    Label-semantic-based prompt tuning for vision transformer adaptation in medical image analysis , author=. IEEE Transactions on Circuits and Systems for Video Technology , year=

  33. [33]

    arXiv preprint arXiv:2506.18378 , year=

    Taming vision-language models for medical image analysis: A comprehensive review , author=. arXiv preprint arXiv:2506.18378 , year=

    Pith/arXiv arXiv

  34. [34]

    The Visual Computer , volume=

    Research progress on AI-assisted screening and prediction of systemic diseases based on retinal images , author=. The Visual Computer , volume=. 2025 , publisher=

    2025

  35. [35]

    Indian Journal of Clinical and Experimental Ophthalmology , volume=

    The role of artificial intelligence in monitoring glaucoma progression using optical coherence tomography , author=. Indian Journal of Clinical and Experimental Ophthalmology , volume=

  36. [36]

    Bioengineering , volume=

    On machine learning in clinical interpretation of retinal diseases using oct images , author=. Bioengineering , volume=. 2023 , publisher=

    2023

  37. [37]

    Eye and Vision , volume=

    Advances and prospects of multi-modal ophthalmic artificial intelligence based on deep learning: a review , author=. Eye and Vision , volume=. 2024 , publisher=

    2024

  38. [38]

    arXiv preprint arXiv:2410.16662 , year=

    Visual Question Answering in Ophthalmology: A Progressive and Practical Perspective , author=. arXiv preprint arXiv:2410.16662 , year=

    Pith/arXiv arXiv

  39. [39]

    The Fourteenth International Conference on Learning Representations , year=

    When large multimodal models confront evolving knowledge: Challenges and explorations , author=. The Fourteenth International Conference on Learning Representations , year=

  40. [40]

    Ophthalmology Science , volume=

    Benchmarking large language models for ophthalmology (BELO): an expert-curated data set and evaluation framework for knowledge and reasoning , author=. Ophthalmology Science , volume=. 2026 , publisher=

    2026

  41. [41]

    arXiv preprint arXiv:2504.11186 , year=

    Benchmarking next-generation reasoning-focused large language models in ophthalmology: a head-to-head evaluation on 5,888 items , author=. arXiv preprint arXiv:2504.11186 , year=

    Pith/arXiv arXiv

  42. [42]

    British Journal of Ophthalmology , volume=

    From visual question answering to intelligent AI agents in ophthalmology , author=. British Journal of Ophthalmology , volume=. 2026 , publisher=

    2026

  43. [43]

    Translational Vision Science & Technology , volume=

    Artificial intelligence for optical coherence tomography in glaucoma , author=. Translational Vision Science & Technology , volume=. 2025 , publisher=

    2025

  44. [44]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    MedBLINK: Probing Visual Perception in Multimodal Language Models for Medicine , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  45. [45]

    Advances in Neural Information Processing Systems , volume=

    DrVD-Bench: Do Vision-Language Models Reason Like Human Doctors in Medical Image Diagnosis? , author=. Advances in Neural Information Processing Systems , volume=

  46. [46]

    Scientific data , volume=

    OCT5k: A dataset of multi-disease and multi-graded annotations for retinal layers , author=. Scientific data , volume=. 2025 , publisher=

    2025

  47. [47]

    2022 international conference on computer communication and informatics (ICCCI) , pages=

    Classification of retinal oct images using deep learning , author=. 2022 international conference on computer communication and informatics (ICCCI) , pages=. 2022 , organization=

    2022

  48. [48]

    Scientific Data , volume=

    Oimhs: An optical coherence tomography image dataset based on macular hole manual segmentation , author=. Scientific Data , volume=. 2023 , publisher=

    2023

  49. [49]

    Retinal OCT Image Classification - C8 , url=

    Obuli Sai Naren , year=. Retinal OCT Image Classification - C8 , url=. doi:10.34740/KAGGLE/DSV/9595300 , publisher=

    arXiv

  50. [50]

    Scientific Data , volume=

    AMD-SD: an optical coherence tomography image dataset for wet AMD lesions segmentation , author=. Scientific Data , volume=. 2024 , publisher=

    2024

  51. [51]

    Scientific data , volume=

    Octdl: Optical coherence tomography dataset for image-based deep learning methods , author=. Scientific data , volume=. 2024 , publisher=

    2024

  52. [52]

    Learning Two-Stream CNN for Multi-Modal Age-Related Macular Degeneration Categorization , year=

    Weisen Wang and Xirong Li and Zhiyan Xu and Weihong Yu and Jianchun Zhao and Dayong Ding and Youxin Chen , journal=. Learning Two-Stream CNN for Multi-Modal Age-Related Macular Degeneration Categorization , year=. 2022 , doi=

    2022

  53. [53]

    International Workshop on Ophthalmic Medical Image Analysis , pages=

    Dataset and evaluation algorithm design for goals challenge , author=. International Workshop on Ophthalmic Medical Image Analysis , pages=. 2022 , organization=

    2022

  54. [54]

    [Zhonghua yan ke za zhi] Chinese journal of ophthalmology , volume=

    Chinese expert consensus on the standardization of OCT and image interpretation and reporting in fundus diseases (2025): a Delphi approach , author=. [Zhonghua yan ke za zhi] Chinese journal of ophthalmology , volume=

    2025

  55. [55]

    Evidence-based guidelines for diagnosis and treatment of diabetic retinopathy in China (2022) , author=. Chin. J. Ocular Fund. Dis , volume=

    2022

  56. [56]

    The Asia-Pacific Journal of Ophthalmology , volume=

    Treat-and-extend regimens for the management of neovascular age-related macular degeneration and polypoidal choroidal vasculopathy: consensus and recommendations from the Asia-Pacific Vitreo-retina Society , author=. The Asia-Pacific Journal of Ophthalmology , volume=. 2021 , publisher=

    2021

  57. [57]

    Comprehensive Adult Medical Eye Evaluation Preferred Practice Pattern

    Wallace, David K and Flaxel, Christina J and Gedde, Steven J and Jacobs, Deborah S and Kopplin, Laura J and Lee, Bryan S and Mah, Francis S and Oetting, Thomas A and Varu, Divya M and Musch, David C , journal=. Comprehensive Adult Medical Eye Evaluation Preferred Practice Pattern

  58. [58]

    Age-related macular degeneration preferred practice pattern

    Vemulakonda, G Atma and Bailey, Steven T and Kim, Stephen J and Kovach, Jaclyn L and Lim, Jennifer I and Ying, Gui-shuang and Flaxel, Christina J and others , journal=. Age-related macular degeneration preferred practice pattern

  59. [59]

    Diabetic retinopathy preferred practice pattern

    Lim, Jennifer I and Kim, Stephen J and Bailey, Steven T and Kovach, Jaclyn L and Vemulakonda, G Atma and Ying, Gui-shuang and Flaxel, Christina J , journal=. Diabetic retinopathy preferred practice pattern. 2025 , publisher=

    2025

  60. [60]

    Idiopathic macular hole preferred practice Pattern

    Kim, Stephen J and Lim, Jennifer I and Bailey, Steven T and Kovach, Jaclyn L and Vemulakonda, G Atma and Ying, Gui-shuang and Flaxel, Christina J and others , journal=. Idiopathic macular hole preferred practice Pattern

  61. [61]

    Retinal and ophthalmic artery occlusions preferred practice pattern

    Kovach, Jaclyn L and Bailey, Steven T and Kim, Stephen J and Lim, Jennifer I and Vemulakonda, G Atma and Ying, Gui-Shuang and Flaxel, Christina J and others , journal=. Retinal and ophthalmic artery occlusions preferred practice pattern

  62. [62]

    2021 , publisher=

    Handbook of Retinal OCT: Optical Coherence Tomography E-Book , author=. 2021 , publisher=

    2021

  63. [63]

    2014 , publisher=

    OCT atlas , author=. 2014 , publisher=

    2014

  64. [64]

    Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    Amo-bench: Large language models still struggle in high school math competitions , author=. Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    2026

  65. [65]

    arXiv preprint arXiv:2604.11778 , year=

    General365: Benchmarking General Reasoning in Large Language Models Across Diverse and Challenging Tasks , author=. arXiv preprint arXiv:2604.11778 , year=

    Pith/arXiv arXiv

  66. [66]

    arXiv preprint arXiv:2510.19316 , year=

    KORE: Enhancing Knowledge Injection for Large Multimodal Models via Knowledge-Oriented Augmentations and Constraints , author=. arXiv preprint arXiv:2510.19316 , year=

    Pith/arXiv arXiv

  67. [67]

    Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    Mined: Probing and updating with multimodal time-sensitive knowledge for large multimodal models , author=. Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    2026

  68. [68]

    arXiv preprint arXiv:2511.02650 , year=

    Can Visual Input Be Compressed? A Visual Token Compression Benchmark for Large Multimodal Models , author=. arXiv preprint arXiv:2511.02650 , year=

    arXiv

  69. [69]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Benchmarking multimodal knowledge conflict for large multimodal models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  70. [70]

    International Conference on Learning Representations , volume=

    Mmke-bench: A multimodal editing benchmark for diverse visual knowledge , author=. International Conference on Learning Representations , volume=

  71. [71]

    International Conference on Learning Representations , volume=

    In-context editing: Learning knowledge from self-induced distributions , author=. International Conference on Learning Representations , volume=