List of Papers Browse by Subject Areas Author List
Abstract
Ocular ultrasound is a critical imaging modality for assessing vitreoretinal pathologies, such as retinal detachment, vitreous hemorrhage, and intraocular tumors. However, interpretation requires specialized expertise and exhibits substantial inter-rater variability, limiting access in resource-constrained settings. Although prior studies have applied deep learning (DL) to ocular ultrasound analysis, progress has been constrained by the lack of public datasets and standardized benchmarks. In this work, we introduce Oculo, a publicly available multi-label dataset comprising 1,630 B-scan images from 1,242 patients, annotated for five ophthalmic abnormalities. We systematically benchmark the performance of four DL backbones and three domain-specific foundation models, across two task settings and three training strategies. We find that EfficientNet-B0 with full-finetuning achieves the best mean performance (F1 = 0.77) across pathology labels. Oculo provides a foundation for standardized evaluation and future methodological advances in automated ocular ultrasound analysis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3020_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Sri-Kanchi-Kamakoti-Medical-Trust/Oculo
Link to the Dataset(s)
Oculo dataset: https://huggingface.co/datasets/SankaraEyeHospital/Oculo
BibTex
@InProceedings{KumSne_Oculo_MICCAI2026,
author = { Kumari, Sneha AND Shah, Payal AND Ganatra, Vaibhav AND Mahesh, Meenakshi AND Jain, Mohit},
title = { { Oculo: A Multilabel Dataset for the Identification of Ocular Abnormalities from Ultrasound Images } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS ActiveX VT_ERROR: },
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces Oculo, a publicly available, multi-label ocular ultrasound dataset comprising 1,630 B-scan images from 1,242 patients, annotated for multiple ophthalmic abnormalities. The authors aim to address the lack of open benchmarks in ocular ultrasound analysis by providing a clinically grounded dataset and systematically evaluating a range of deep learning architectures and foundation models. The paper benchmarks four standard backbones (EfficientNet, ResNet, VGG, ViT) under both single-task and multi-task learning settings, and further evaluates three domain-specific foundation models (VisionFM, USFM, OpenUS) across different fine-tuning strategies. Overall, the work focuses on dataset construction and benchmarking rather than proposing new methods, and aims to establish a standardized evaluation resource for ocular ultrasound research.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper introduces the first publicly available, multi-label ocular ultrasound dataset annotated for multiple vitreoretinal and choroidal abnormalities. This addresses a significant gap in the field. 2.The authors provide a systematic comparison of conventional ML methods and foundation models. The discussion on texture vs. shape bias and the observed degradation under full fine-tuning for some foundation models provides useful insights beyond raw performance numbers.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The paper does not introduce new modeling or learning strategies. But this would not affect my justification because the main contribution of this paper is the dataset. 2.Several rare pathologies are excluded from evaluation due to limited sample sizes. 3.The dataset is collected from a single center using a single ultrasound device. 4.Only static B-scan images are considered, whereas ocular ultrasound is often acquired as video sequences.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
The proposed dataset contains private information. However, the authors do not show any ethical approval document from IRB or a clinical trial registration number.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Any new dataset could be a treasure for the community. However, the authors should prove they have already acquired ethical approval from IRB.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Though there are several formatting or typo problems. The contribution of the new dataset is always welcome.
Review #2
- Please describe the contribution of the paper
The paper introduces Oculo, a publicly available multi-label dataset of 1,630 ocular ultrasound B-scan images from 1,242 patients, annotated for eight ophthalmic abnormalities (five used in experiments). The authors benchmark four DL architectures (EfficientNet-B0, ResNet50, VGG-19-BN, ViT-B-16) and three domain-specific foundation models (VisionFM, USFM, OpenUS) across single-task and multi-task settings with three training strategies. Key findings: EfficientNet-B0 with full fine-tuning achieves best mean F1 (0.77 multi-task), ViT improves with multi-task training while CNNs degrade, and foundation model linear probes nearly match fully fine-tuned EfficientNet.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.This is the first public multi-label ocular ultrasound dataset. Table 1 clearly shows that all prior work used private data. Releasing this dataset addresses a genuine bottleneck in the field and enables reproducible evaluation.
2.The annotation protocol is well-designed: two senior ophthalmologists independently annotated after a calibration phase, with a retina specialist (16 years experience) adjudicating ambiguous cases. This is a rigorous process for establishing ground truth.
3.The benchmarking is systematic and covers a useful range of comparisons: CNN vs. transformer, single-task vs. multi-task, full fine-tuning vs. linear probing, and ImageNet-pretrained vs. domain-specific foundation models. The observation that ViT improves in multi-task while CNNs degrade, and the texture/shape bias interpretation (Fig. 2), provides actionable insight.
4.The foundation model analysis is timely and reveals a practical finding: USFM and VisionFM linear probes (F1: 0.76) nearly match fully fine-tuned EfficientNet-B0 (F1: 0.77). This suggests foundation models can be effective without task-specific fine-tuning, which is important for resource-constrained settings.
5.Patient-level splits prevent data leakage. 5-run averages are reported. Code and data will be released.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.No inter-rater agreement metrics are reported. The paper describes two independent annotators with specialist adjudication, but does not report Cohen’s kappa, Fleiss’ kappa, or any agreement statistic. For a dataset paper, this is a significant omission because it is impossible to assess label reliability without it.
2.Three of eight annotated labels (PVD, CD, PB) are excluded from all experiments due to low prevalence (17, 29, 19 cases). This reduces the effective dataset to five labels. For a paper whose title emphasizes “multilabel” identification, losing 37.5% of labels weakens the contribution. Furthermore, it limits the dataset’s utility for studying rare conditions, which is where automated tools would be most valuable.
3.Single device, single institution. All images are from one ultrasound system (Appasamy Marvel II, 12 MHz) at one hospital. Device-specific image characteristics (resolution, gain settings, speckle patterns) could bias learned features. The paper does not discuss generalization to other devices or clinical settings.
4.No methodological contribution. The paper benchmarks existing architectures and foundation models with standard training protocols. While this is appropriate for a dataset paper, the benchmarking itself does not introduce any new technique or insight beyond what would be expected from a straightforward application of known methods to a new dataset.
5.The multi-task performance drop for CNNs (EfficientNet F1: 0.92 single-task to 0.77 multi-task) is noted but not analyzed in depth. A shared backbone with separate heads is the simplest multi-task design. More sophisticated approaches (task-specific feature adaptation, loss weighting strategies, gradient balancing) are not explored.
6.No comparison with the existing ERDES dataset [14]. Since ERDES contains retinal detachment annotations, a cross-dataset evaluation on this overlapping label would strengthen the paper’s contribution.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Dataset papers at MICCAI are judged primarily on (1) the significance of the gap they fill, (2) dataset quality and annotation rigor, and (3) the usefulness of the benchmarks provided.
On (1), the gap is clear: no public multi-label ocular ultrasound dataset existed before. This limits reproducibility across the entire subfield, as Table 1 demonstrates. On (2), the annotation protocol involves multiple experts with calibration and adjudication, though the lack of inter-rater agreement metrics is a notable gap. On (3), the benchmarking is systematic and provides useful insights about architecture choices and foundation model transferability.
The main weaknesses are the missing inter-rater statistics, single-device/single-institution limitation, exclusion of three labels, and absence of methodological novelty. However, releasing a public dataset in an underserved domain has outsized community impact. If the authors report inter-rater agreement and discuss device generalization limitations in the rebuttal, the paper would be strengthened.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper introduces Oculo, a multi-label ocular ultrasound (B-scan) dataset annotated for eight ophthalmic abnormalities, comprising 1,630 images from 1,242 patients. Alongside the dataset, the authors provide a systematic benchmark of four deep learning backbone architectures (EfficientNet-B0, ResNet50, VGG-19-BN, and ViT-B-16) and three domain-specific foundation models (VisionFM, USFM, and OpenUS), evaluated on five primary labels across single-task and multi-task settings under different fine-tuning settings.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The dataset is claimed as the central contribution of this work. Considering the scarcity of publicly available ophthalmic ultrasound data, the open release of this dataset would make a valuable contribution to the community.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The manuscript exhibits several formatting issues that are clearly inconsistent with the MICCAI template, including the title and all tables (e.g., the placement of table titles, alignment, and spacing). Noncompliance with the MICCAI formatting requirements is unacceptable. 2.The manuscript contains multiple inconsistencies in presentation, suggesting a lack of careful proofreading. The most notable example is in Table 3, where the specificity of VGG-19-BN is reported as 0.98, whereas all other entries in the same column are presented as percentages, indicating inconsistent numerical formatting. Other examples include inconsistent terminology such as “full finetuning” versus “full-finetuning”. 3.In Table 4, the performance of full-parameter fine-tuning is generally lower than that of linear probing, which is contrary to expectations. The authors attribute this to domain shift and differences in image resolution arising from varying ultrasound acquisition settings. This explanation is not convincing, as full fine-tuning is theoretically expected to better accommodate domain differences.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The manuscript exhibits clear noncompliance with the MICCAI formatting requirements, as well as noticeable errors and inconsistencies that suggest a lack of careful proofreading. While the public release of the dataset would provide a valuable contribution to the community, the authors do not clarify under what conditions the dataset will be made available.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
Although the dataset promised to be publicly available in this paper has practical significance, the numerous formatting issues and input errors repeatedly mentioned in the paper have undermined its credibility (such as the inconsistent decimal places in Table 2 and the inconsistent unit of indicators in Table 3). Furthermore, the issues regarding the inconsistent experimental results of linear probing and full fine-tuning have not been satisfactorily explained. The authors should disclose more experimental details to support the rationality of the relevant experimental results.
Author Feedback
We thank the reviewers for their thoughtful comments. We are glad that all reviewers appreciate the value of our ocular ultrasound dataset.
# ETHICS APPROVAL (R1,M2) Informed consent was obtained from all patients as part of our study protocol, which was approved by the hospital’s IRB (Sec. 2). The IRB number was omitted to preserve anonymity and will be included in the camera-ready version. The IRB approval also explicitly covers public release of the dataset. Furthermore, all personal health information has been fully de-identified.
# DATASET + CODE RELEASE (R1,R4,M2) We will publicly release the dataset and code under MIT License upon acceptance of the paper.
# INTER-RATER AGREEMENT (R2,M2) Two ophthalmologists independently annotated all 1,630 images, with disagreements adjudicated by a senior retina specialist (Sec. 2). We computed Cohen’s Kappa (K) for each label, and the average K for all labels was 0.86 with an agreement of 96.7%. This confirms reliable ground truth quality across labels.
# SINGLE CENTER AND DEVICE LIMITATIONS (R2,M2) Generalization across devices and sites remains a long-standing challenge in medical imaging. Even in our dataset, being single center and single device is a limitation. We hope that our public dataset will enable future studies evaluating generalization across datasets from different sites, populations, and devices.
# EXCLUSION OF RARE LABELS (R1,R2,M2) Throughout the paper (including the Abstract and Introduction), we claim annotations for five ophthalmic abnormalities as the core contribution. To the best of our knowledge, no publicly available ocular ultrasound dataset currently provides labels for these five abnormalities. The three additional labels are described in the annotation section for completeness; however, due to limited data availability, they were not included in our experimental evaluation. Given the rarity of these conditions, we believe that even a modest number of annotated examples has clinical and research value, and we therefore chose to release them as supplementary samples. We clarify that this is not a case of overclaiming, as we do not present all eight abnormalities as our core contribution. To make this distinction clearer, we will gray out the corresponding three entries in Table 2.Finally, we are continuing data collection for these conditions, and future versions of the dataset will include additional samples.
# FORMATTING ERRORS AND TYPOS (R4,M2) We apologise for placing the table captions below the table and will fix it in the camera-ready version. Also, we will proofread the paper and fix all typos (including in Table 3).
# FINETUNING VS LINEAR PROBING (R4,M2) The lower finetuning performance compared to linear probes can be attributed to a combination of domain shift, class imbalance, and modest dataset size relative to the foundation model size [1, Paper Ref# 11]. Notably, Ruffini et al. [1] found that finetuning biomedical foundation models on small, imbalanced datasets can underperform compared to linear probing. Our results align with these findings.
# STATIC B-SCAN IMAGES VS VIDEOS (R1) We agree that ocular ultrasound is captured as video sequences; however, we chose to focus on static frames. In a typical ultrasound video, only a subset of frames exhibit pathology, and even abnormal scans often contain many normal frames. This creates significant class imbalance at the frame level, making model training challenging. By manually extracting and annotating individual frames, we provide a complete characterization of the pathologies present in each frame. Also, it reduces the annotation burden on ophthalmologists.
# NO METHODOLOGICAL CONTRIBUTION (R2) As R1 and R2 noted, a methodological contribution is not expected from a dataset paper.
[1] Ruffini, F et al. (2025). Benchmarking foundation models and parameter-efficient fine-tuning for prognosis prediction in medical imaging. Computer Methods and Programs in Biomedicine, 109196.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper introduces Oculo, a public multi-label ocular ultrasound dataset with benchmarking of standard deep learning architectures and foundation models. Reviewers agree that the dataset addresses an important gap in ocular ultrasound research and could provide useful value to the community.
However, several concerns need clarification. The dataset is single-center and single-device, three of the eight annotated labels are excluded from the experiments due to low prevalence, and no inter-rater agreement statistics are reported despite the expert annotation protocol. Reviewers also raised concerns about formatting/template compliance, numerical inconsistencies, and the need to clarify the exact conditions for public data/code release and ethics approval.
The authors should address: 1- Ethics/IRB approval and data release conditions. 2- Inter-rater agreement statistics. 3- Single-center/single-device limitations and generalization. 4- Exclusion of rare labels and implications for the “multi-label” claim. 5- Formatting/template issues and numerical inconsistencies. 6- Clarification of unexpected foundation model results, especially linear probing vs. full fine-tuning.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The main pre-rebuttal concerns were reasonable but mostly addressable: missing IRB details, lack of inter-rater agreement statistics, single-center/single-device limitation, exclusion of three rare labels from experiments, and several formatting/presentation issues. In rebuttal, the authors clarified that IRB approval and informed consent were obtained, that dataset release is covered by the approval, that the data will be publicly released with code, and they provided inter-rater agreement statistics with strong agreement. They also clarified the role of the three rare labels and acknowledged the single-center limitation as a limitation rather than an overclaim.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After considering the reviews and the rebuttal, I recommend acceptance. The paper presents a public multi-label ocular ultrasound dataset and a useful benchmark for an underrepresented imaging modality. The rebuttal clarifies the IRB approval, consent, dataset/code release plan, inter-rater agreement, rare-label handling, and single-center/device limitation. It also explains the linear probing versus fine-tuning results and commits to correcting formatting and presentation issues. I therefore recommend acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper is recommended for acceptance because it introduces the first publicly available, multi-label ocular ultrasound dataset, filling a major resource gap in the community. Reviewers commended the rigour of the annotation protocol, which utilised multiple independent experts and senior adjudication. The authors have also committed to correcting all formatting and typographical errors in the final template. Given the outsized community impact of releasing an open benchmark in an underserved domain, the paper meets the publication threshold.