List of Papers Browse by Subject Areas Author List
Abstract
Spherical Equivalent Refraction (SER) and Axial Length (AL) are core indicators for pediatric myopia screening, yet their measurements require dedicated biometry and cycloplegic refraction. Fundus photography offers an accessible imaging modality, as myopia-related posterior-pole changes are visible in 45◦ fundus images. However, these cues are often low-contrast, spatially diffuse, and multi-scale. Moreover, AL, Sphere (SPH), and Cylinder (CYL) share partially overlapping but non-identical anatomical correlates. We propose SpecF2M, a spectral-aware multi-task network for estimating AL and SER components from pediatric fundus photographs. SpecF2M integrates a deterministic anatomy-guided enhancement module, a hybrid spatial–spectral backbone combining MixCNN and Hybrid Spectral Learning (HSL) blocks, and an expert-routing head for component-level estimation of AL, SPH, and CYL. On a pediatric cohort of 4,359 eligible child visits and 6,966 fundus images, SpecF2M outperforms controlled CNN/ViT baselines for AL and SPH estimation, achieving MAEs of 0.5347 mm and 0.7062 D, respectively. Component-level analysis further reveals asymmetric task coupling, where CYL exhibits weaker association with fundus-derived myopic patterns than AL/SPH. These results support fundus-based, screening-oriented estimation of pediatric myopia indicators, while external validation remains necessary before deployment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1928_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HeMen_SpecF2M_MICCAI2026,
author = { He, Mengxian AND Liu, Xinyue AND Sun, Yunyun AND Wei, Hao AND Zhang, Minqing AND Zhang, Shunyi AND Wang, Lichun AND Yuan, Wu},
title = { { SpecF2M: A Spectral-Aware Multi-task Network Estimating Axial Length and Refractive Error from Pediatric Fundus Photographs } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes SpecF2M, a deep learning framework designed to jointly estimate Axial Length (AL) and refractive error components (Sphere [SPH] and Cylinder [CYL]) from 45° pediatric fundus photographs. The pipeline integrates a deterministic anatomy-guided image enhancement module, a hybrid backbone combining spatial (MixCNN) and spectral (Wavelet/Fourier) feature learning, and an expert-routing multi-task head to output AL, SPH, and CYL simultaneously.
- 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.
High Clinical Relevance: Targeting the joint estimation of AL, SPH, and CYL from widely accessible 45° fundus images directly addresses a major logistical bottleneck in large-scale pediatric myopia screening.
Solid Architectural Engineering: The hybrid feature extraction strategy is structurally sound. By combining deformable convolutions with multi-resolution wavelet and Fourier transforms, the network is well-equipped to capture the multi-scale, spatially diffused, and geometric deformations characteristic of myopic fundus changes.
Promising Internal Performance: Evaluated on a relatively large clinical cohort (N=4,359), the model demonstrates strong predictive capabilities for AL (MAE: 0.5347 mm) and SPH (MAE: 0.7062 D), outperforming several standard deep learning baselines.
- 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.Lack of Dataset Diversity and External Validation: The evaluation is strictly confined to a single cohort (N=4,359) with a simple 80/10/10 internal split based on subject IDs. Relying on a single-center, single-device data source is entirely insufficient to demonstrate the model’s generalizability across diverse demographics, varying fundus camera brands, or different imaging conditions. Consequently, the claim that SpecF2M is a “scalable” tool for large-scale screening is unsubstantiated without an independent external test set. 2.Outdated Quantitative Baselines: The quantitative benchmark relies on antiquated general-purpose architectures (ResNet18, DenseNet121, VGG16, and standard ViT). The experimental design completely omits comparisons with recent, state-of-the-art deep learning models specifically designed for ophthalmic imaging or multi-task medical image analysis, which severely undermines the credibility of the performance claims. 3.Weak and Insufficient Qualitative / Ablation Experiments: The experimental design lacks depth in validating the proposed mechanisms: 1) Interpretability (Grad-CAM): The qualitative evaluation relies solely on standard Grad-CAM visualizations compared only against a basic ResNet18.The authors fail to visualize the high/low-frequency feature maps generated by the WSConv or WPO modules, leaving the actual mechanism and efficacy of the spectral learning unproven.
- 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 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the paper addresses a valuable clinical need and achieves decent internal performance, its claims of real-world scalability are critically undermined by the complete lack of external validation on independent datasets. Furthermore, the experimental design is flawed, relying on outdated baselines (e.g., ResNet18, VGG16) and superficial qualitative analyses that fail to rigorously validate the internal mechanisms of the proposed spectral modules. Finally, the methodology overstates the novelty of standard, non-learnable preprocessing operators and misses a vital opportunity to enforce true multi-task disentanglement by failing to incorporate the established physiological constraint ($SER = SPH + 0.5 \times CYL$) into the loss function.
- 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.
N/A
Review #2
- Please describe the contribution of the paper
SpecF2M is a multitask regression framework that jointly estimates Axial Length (AL), Sphere (SPH), and Cylinder (CYL) from 45° pediatric fundus photographs. The core novelty is threefold: a training-free anatomy-guided enhancement pipeline targeting myopia-specific fundus structures (OD, PPA, fundus tessellation, vessels); a spectral-aware backbone interleaving wavelet-selective CNN layers with Fourier-domain wave propagation operators for multi-scale feature learning; and an expert-routing multitask head that disentangles shared versus task-specific representations across AL, SPH, and CYL. The clinical premise is replacing dedicated biometry devices and cycloplegic refraction with a single fundus photograph for large-scale pediatric myopia screening.
- 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.Clinically relevant and scalable problem. Pediatric myopia screening at population scale is a genuine public health need, particularly in East Asia. Replacing Lenstar biometry and cycloplegic refraction with a fundus photograph is a meaningful clinical goal with real deployment potential in school screening programs. 2.Dataset 4,359 children with paired fundus, AL, and cycloplegic refraction data is a strong dataset for this task. The subject-level split to prevent data leakage is correctly implemented and explicitly stated. This is a notable strength compared to most papers in this space. 3.The anatomy-guided enhancement is well-motivated and training-free. The cascade of CLAHE, high-frequency emphasis, vessel suppression, and OD/PPA ROI enhancement is grounded in the specific visual characteristics of myopic fundus changes. Being training-free means it adds no parameters and generalizes without overfitting risk. The physiological rationale for each step is clear. 4.Ablation is honest and informative. Table 1 combines baselines and ablations in a single table. The ablation correctly identifies that CYL benefits less from the framework than AL and SPH, and the authors do not hide this — they explain it via the correlation analysis in Fig. 3 showing AL-CYL near-zero correlation. This is scientifically honest and adds mechanistic insight. 5.The asymmetric coupling finding is genuinely interesting. The observation that AL and SPH share anatomical correlates (corr = -0.521) while CYL does not (corr = -0.068) is a clinically meaningful insight that goes beyond performance numbers. This has implications for how multitask myopia screening systems should be designed. 6.R² values are substantially more credible than comparable work. SPH R² = 0.73 and AL R² = 0.60 on a real-world pediatric cohort are competitive results. Unlike the previous paper reviewed, the performance here is actually in a range where clinical utility is plausible for a screening tool.
- 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 external validation. The training-free enhancement pipeline helps but does not substitute for cross-center validation. For a tool proposed for large-scale population screening this is a significant gap. 2.The HSL block combines two spectral components without sufficient justification. WSConv (wavelet) and WPO (Fourier wave propagation) are cascaded but their individual contributions are not ablated separately. The w/o HSL block ablation removes both simultaneously. The reader cannot determine whether the gain comes from wavelet decomposition, wave propagation, or their combination. This is an incomplete ablation for what is presented as the core technical contribution. 3.The Wave Propagation Operator is borrowed with minimal adaptation. WPO is taken directly from Waveformer (Shu et al., arXiv 2026 notably a 2026 arXiv preprint cited as reference [18]). The physical motivation for applying underdamped wave dynamics to fundus feature maps is not adequately explained beyond a reference to the original paper. Why does wave propagation specifically benefit myopic fundus representation over standard global attention or frequency filtering? 4.No comparison against prior AL or SER estimation methods. The paper compares only against generic CNN/ViT backbones in single-task and hard multitask settings. There are published methods specifically for AL estimation from fundus (Dong et al. [1], Jeong et al. [11], Yang et al. [23]) and joint AL+SER estimation (Oh et al. [16], Zhong et al. [25], Li et al. [14]) that are cited in the introduction but not compared against in Table 1.This is not acceptable. If those methods cannot be reproduced exactly, the authors should at minimum compare against their reported numbers on overlapping metrics. 5.Enhancement hyperparameters are fixed and not validated. The anatomy-guided enhancement uses fixed hyperparameters (α for sharpening, λ for vessel suppression, σ for Gaussian, τ for ROI blending). No sensitivity analysis is provided for any of these. For a training-free pipeline that processes every input image, robustness to hyperparameter choice matters significantly, especially across different camera models and image qualities. 6.Confidence intervals absent throughout. No confidence intervals or standard deviations on any metric in Table 1.With a 10% test split of 4,359 subjects (~436 eyes, but this could be fewer unique subjects depending on bilateral inclusion), variance estimates matter.
- 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?
The primary factors driving the score to weak accept are the dataset size, the honest ablation, and the clinical relevance. The factors preventing a stronger accept are the absence of any prior task-specific baselines in Table 1 which is the most damaging gap and the incomplete spectral ablation. If the authors can demonstrate in rebuttal that their method outperforms Dong et al., Yang et al., and Zhong et al. on overlapping metrics, the score would move to accept. The rebuttal should focus entirely on this point.
- 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.
The rebuttal strengthens the submission. The demographic breakdown (age 8.22±1.09y, AL 23.88±1.07mm) and confirmation of subject-level splitting remove two ambiguities I flagged. The commitment to report SER explicitly, add bootstrap CIs, and soften scalability claims addresses the clinical utility framing concern directly. The contextual comparison against Varadarajan (MAE 0.56D, n=60K), Dong (0.56mm), and Oh (0.744mm) is honest, the authors correctly acknowledge direct comparison is confounded by modality and population differences rather than overclaiming SOTA. This is the right scientific posture. The HSL attribution explanation is acceptable. Claiming integrated block effectiveness rather than individual operator attribution is a defensible scope limitation, not a flaw, provided it is stated clearly in the paper. The missing external validation remains a real limitation but is now explicitly acknowledged as future work. Given the 7-school cohort diversity and stratified splits, the internal evaluation is more robust than initially appeared.
Review #3
- Please describe the contribution of the paper
This paper investigates the problem of estimating axial length (AL) and refractive components using widely available 45° pediatric fundus photographs. A multi-task framework, SpecF2M, is proposed for jointly estimating AL, spherical power (SPH), and cylindrical power (CYL) from fundus images. This method comprises three main components: an anatomically based enhancement strategy to highlight relevant myopic patterns; a spectral-aware backbone network combining wavelet decomposition and Fourier domain propagation; and an expert-routing multi-task design to model shared and task-specific factors of AL, SPH, and CYL. Experiments were conducted in a real-world pediatric cohort of 4359 children, and results show that the proposed model achieves state-of-the-art performance among the listed baseline models in terms of AL and SPH estimation.
- 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 paper explores pediatric myopia screening, where axial length (AL) and cycloplegic refraction are crucial but require specialized equipment or good patient cooperation. Exploring whether fundus photography can serve as a more convenient alternative method is of significant importance and practical application value.
2.The proposed design aligns with known physiological priors, namely that myopia-related fundus changes are diffuse, low-contrast, and multi-scale. Anatomically guided enhancement, spectral perceptual representation learning, and multi-task expert routing constitute a coherent overall framework, rather than an arbitrary combination of modules.
3.This paper does not merely directly predict the equivalent spherical power (SER), but explicitly and jointly models axial length (AL), spherical power (SPH), and cylindrical power (CYL). This component-level modeling is not only useful for prediction but also has potential value in analyzing cross-task relationships, and the correlation analysis in this paper provides some support for this view.
- 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.Although this paper focuses on the estimation of SER and AL, the main experimental results only report SPH, CYL, and AL, not SER itself. Since SER is emphasized as a key clinical indicator in the motivation section, the lack of direct assessment of SER is a significant oversight. Since SER can be calculated from the predicted SPH and CYL, it should also be easy to report the MAE and R² of SER.
2.Experimental comparisons primarily included general backbone networks such as ResNet18, DenseNet121, VGG16-BN, and ViT variants. While these are reasonable reference models, they are insufficient to demonstrate best performance for this task. Previous work on AL estimation and joint estimation based on fundus images was discussed, but stronger, task-specific baseline models were not directly compared in the experiments. This weakens its empirical assertions of superiority.
3.The paper states that the cohort included 4359 children, grouped by subject ID, which is satisfactory. However, many important details remain missing, such as the total number of images, number of visits per subject, distribution of AL/SPH/CYL values, age range, device consistency, image quality control, and exclusion criteria. These details are crucial for understanding the task difficulty and the generalizability of the assessment results.
4.This paper only reports MAE and R². For a clinically meaningful screening task, additional analyses would significantly enhance the persuasiveness of the paper, such as subgroup analyses by refractive range or age group, error distribution plotting, Bland-Altman analysis, and evaluation of threshold-based screening performance. Without these analyses, conclusions regarding screening utility remain preliminary.
5.One of the paper’s claimed contributions is the joint estimation of AL, SPH, and CYL. However, the CYL results are less convincing: in Table 1, the proposed method is not the optimal approach for CYL, and the authors themselves point out that the ablation trend is noisy and unstable. This makes the three-task model less convincing than described in the paper.
6.The paper presents an overall framework and some equations, but some key implementation details are missing or insufficiently described, including the specific regression loss used, the selection of multi-task weights, the wavelet type and decomposition depth, the architecture of the expert MLP, data augmentation strategies, and details of the optimization process. These require further clarification.
7.There are some minor language and formatting issues, including grammatical errors, awkward wording, and spelling mistakes. Furthermore, the visual effect of Figure 2 is too dense and difficult to read at a small size, which reduces its effectiveness as a primary method diagram.
- 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 provide sufficient information for 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper explores a clinically significant problem with a well-designed overall framework. Multi-task modeling and spectrum-aware modeling are innovative, and improvements are made in AL and SPH. However, the current manuscript does not provide sufficient and complete evidence to support its main arguments. In particular, the lack of direct SER assessment, limited baseline strength, incomplete dataset description, and relatively weak clinical applicability validation all reduce its overall impact. I suggest the authors strengthen the experimental section and make the assessment more aligned with the established clinical goals.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
I appreciate the authors’ detailed response and their willingness to revise the manuscript. The rebuttal clarifies several points, including the dataset size, exclusion criteria, implementation details, and the rationale for separating SPH/CYL from SER. However, my main concerns are only partially addressed. In particular, SER is emphasized as a key clinical endpoint, but the rebuttal does not provide the actual SER results. The comparison with stronger task-specific baselines remains limited, and the authors mainly propose to soften the claims rather than provide additional controlled comparisons. Moreover, the clinical screening utility is still insufficiently supported by MAE and R² alone, without subgroup analysis, Bland-Altman analysis, error distribution, or threshold-based screening metrics. The weaker and unstable CYL performance also suggests that the claimed joint estimation of AL, SPH, and CYL should be interpreted more cautiously.
Author Feedback
We thank the reviewers for their constructive comments and for recognizing the clinical relevance of our study [Meta,R2,R3,R6], its physiological motivation [R2,R3,R6], the strong AL/SPH performance [Meta,R3,R6], and the scale of the cohort [Meta,R3].
1.Dataset, Ethics, and Generalizability [Meta,R2,R3,R6] We apologize for the imprecise cohort-unit wording: the dataset contains 4,359 eligible child-visits. This does not affect the reported results since splitting was subject level. Following the exclusion of poor-quality images (unclear optic disc, reflection, artifacts), 6,966 fundus images were included. Mean values for age, AL, SPH, and CYL were 8.22±1.09 years, 23.88±1.07mm, -0.19±1.85D, and -0.83±0.95D, respectively. The ethics approval will be disclosed in the camera-ready version without breaking anonymity.
We agree that external validation is needed. Although not external, the current cohort was collected from 7 primary schools and includes both disc- and macula-centered imaging settings. We stratified the validation and test sets according to myopia levels to reduce internal evaluation bias. We will soften scalability claims and clearly state external validation as a limitation.
2.Baselines and experiments [Meta,R2,R3,R6] Direct comparison with task-specific studies is limited by differences in modalities, outputs, and unavailable code/data. For context only, we cite the strongest reported task-specific baselines: for SER estimation, Varadarajan et al. reported MAEs of 0.56D (n=60,127) and 0.91D (n=4,628) using fundus images, while Zhong et al. reported 1.622D using ultrawide-field fundus (UWF). For AL estimation, Dong et al. reported 0.56mm using fundus images, and Oh et al. reported 0.744mm using UWF. Since these studies generally do not report separate SPH/CYL results, we will restrict claims to controlled internal comparisons and avoid overclaiming SOTA.
3.SER, CYL, and loss design [R2,R3,R6] We agree that SER is clinically important, but decomposing it into SPH and CYL can exposes component-level behaviours hidden by SER-only estimation. Since e_SER=e_SPH+0.5×e_CYL, opposite SPH/CYL residuals can partially cancel and make SER error appear lower while masking component-specific failures. We will clarify this rationale, report SER and add residual analysis to quantify SPH/CYL contributions to SER error, and report Bootstrap CIs and repeat-run SDs for metric stability.
CYL is less coupled with posterior-pole fundus structures, as astigmatism is often cornea-driven. Its weaker AL–CYL correlation and less stable performance suggest higher noise. Therefore, SER was not used as an auxiliary loss, since it is linearly redundant with SPH/CYL and may bias learning toward the aggregate target.
4.Spectral/HSL [R2,R3,R6] We appreciate the request for finer WSConv-vs-WPO attribution. HSL is designed as a coupled spectral block: WSConv extracts local multi-resolution subband cues, while WPO propagates global frequency-domain information. In our initial experiment, variants removing WSConv or WPO showed degradation. Thus, our primary claim is the effectiveness of integrated HSL for AL/SPH estimation, not complete causal attribution of each operator. Physiologically, myopic fundus cues are diffuse and multi-scale, motivating the combination of local subband modeling and global spectral propagation. We will clarify this claim boundary and discuss finer attribution/visualization as follow-up analyses.
5.Implementation details [R6] For training, the submitted experiments used L2/MSE regression loss, multi-task weights of AL: SPH: CYL=0.5:1:1, and only geometric augmentations, including random rotation, flipping, and mirroring. For the model, the WSConv applied the Daubechies-1 wavelet and 2-level decomposition. The expert MLPs use a one-layer 512-d MLP with ReLU and dropout=0.2. 6.Writing typo[R6] We will also correct formatting issues and improve figure readability in the revision.
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.
This paper proposes a multi-task framework for estimating axial length and refractive components from pediatric fundus photographs, with the goal of supporting large-scale myopia screening in children. The problem is clinically meaningful, and the study is supported by a relatively large dataset. However, the current submission still has several important limitations. In particular, the evaluation lacks external validation, does not include stronger task-specific baselines, and does not yet provide sufficiently complete evidence for the contribution of the proposed spectral components. In addition, important dataset, implementation, and clinical evaluation details remain insufficiently described. The manuscript also uses a pediatric clinical cohort but does not clearly report ethical approval or informed consent, which should be clarified. The authors are encouraged to address these concerns in the rebuttal.
- 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 addresses a clinically relevant problem and presents a coherent framework for estimating AL, SPH, and CYL from pediatric fundus photographs. The rebuttal clarifies the cohort details, subject-level splitting, ethics approval, implementation details, and the scope of the comparison with prior task-specific studies. It also clarifies the role of SER, confidence intervals, and the intended scope of the results. I therefore recommend acceptance.
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 reviewing the manuscript, the reviews, and the authors’ rebuttal, I recommend acceptance.
The rebuttal adequately addresses most of the major concerns raised during the review process, particularly regarding cohort characteristics, data splitting, clinical framing, and the positioning of the work relative to prior studies. The study is based on a sizeable cohort and presents a technically sound approach to a clinically relevant problem.
I acknowledge the remaining concerns regarding external validation and the lack of additional clinical analyses such as subgroup evaluation and screening-oriented metrics. These would strengthen the paper and should be considered in future work. However, I view them as limitations of validation scope rather than fundamental weaknesses of the methodology or experimental design.
Overall, the paper makes a useful contribution and the current evidence is sufficient to support its main claims. Therefore, I recommend acceptance.