Authors

Zhen Yu, Ruiye Chen, Peng Gui, Lie Ju, Xianwen Shang, Zhuoting Zhu, Mingguang He, Zongyuan Ge

Abstract

Retinal age has recently emerged as a reliable ageing biomarker for assessing risks of ageing-related diseases. Several studies propose to train deep learning models to estimate retinal age from fungus images. However, the limitation of these studies lies in 1) both of them only train models on snapshot images from single cohorts; 2) they ignore label ambiguity and individual variance in the modeling part. In this study, we propose a progressive label distribution learning (LDL) method with temporal fundus images to improve the retinal age estimation on snapshot fundus images from multiple cohorts. First, we design a two-stage LDL regression head to estimate adaptive age distribution for individual images. Then, we eliminate cohort variance by introducing a domain-aware ordinal constraint to align image features from distinct data sources. Finally, we add a temporal branch to model sequential fundus images and use the captured temporal evolution as auxiliary knowledge to enhance the model’s predictive performance on snapshot fundus images. We use a large retinal fundus image dataset which consists of around 130k images from multiple cohorts to verify our method. Extensive experiments provide evidence that our model can achieve lower age prediction errors than existing methods.

Link to paper

DOI: https://doi.org/10.1007/978-3-031-43990-2_59

SharedIt: https://rdcu.be/dnwMe

Link to the code repository

N/A

Link to the dataset(s)

N/A

Reviews

Review #2

Please describe the contribution of the paper

A framework to estimate retinal age from fundus images is proposed. A two-stage strategy is adopted. In addition, domain-aware ordinal constraints and temporal branch are introduced to improve network performance.
Please list the main strengths of the paper; you should write about 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.

Label distribution learning, domain-aware ordinal constraints and temporal data are employed to address shortcomings in previous work . Domain-aware ordinal constraints are introduced to improve performance in different data source. It’s a new application of these algorithm in field of retinal age estimation. Experiments are clear and well designed.
Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

The definition of Qcoarse in section 2.1 is not clear enough. As one of the main contributions of the paper, section 2.3 didn’t provide enough detail. What is the architecture of time dimensional attention module (TDA)? How to compute the distance correlation? This paper only gave a reference without detailed description. Multiple losses only add up to final loss without corresponding investigation.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance

Reproducibility of the results reported in this paper is moderate. The author didn’t provide source code yet and dataset will not be available. The framework contains three main parts and the detail definition in section 2.3 just gives a reference, which decreases the reproducibility.
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html

The spelling of “fundus” is wrong in abstract. The definition of Qcoarse in section 2.1 is not clear enough. Give more details in section 2.3. Did you considered the computational complexity or run time? PLDL_temp in the table is not described. The improvement between PLDL and PLDL_temp is very limited.
Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making

5
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

A framework is proposed to estimate retinal age from fundus images. The author applied label distribution learning, domain-aware ordinal constraints and temporal data to improve the performance. The experiments show the improvement than previous work. However, some methods are already common processing algorithms and the paper only did a novel application. Part of the expression of the paper still needs to be improved.
Reviewer confidence

Confident but not absolutely certain
[Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

N/A
[Post rebuttal] Please justify your decision

N/A

Review #3

Please describe the contribution of the paper

This study proposes a new method called “progressive label distribution learning (LDL)” to improve retinal age estimation from temporal fundus images. The method consists of three key steps: estimating adaptive age distribution, eliminating data source variance, and adding a temporal branch to model the temporal properties of fundus images.
Please list the main strengths of the paper; you should write about 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) using a large-scale dataset with ~130k images from multiple cohorts; 2) formulating the age estimation as a two-stage label distribution learning task and giving an adaptive distribution estimate for individual images; 3) introducing domain-aware ordinal constraints to align the image features from different data sources; 4) adding a temporal branch to leverage the auxiliary knowledge from temporal data to enhance the performance on snapshot images.
Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

1) The paper could provide more details on the dataset and implementation, such as the selection criteria for healthy subjects, the number and distribution of age labels, and the specific data augmentation techniques used. 2) lack of evaluation on different age groups and subgroups; 3) lack of discussion on the limitations and future directions of the proposed method.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance

The paper provides sufficient details on the methodology and experimental setup to allow for reproducibility.
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html

1) The order of references is confusing, why not start with 1? 2) Spelling errors, e.g. in the abstract “from fungus images.” shoud be “from fundus images.” 3) Do L_1and L_2 in Fig.1 correspond to L_lds? 4) In Ablation study,” Table 5b” shoud be Fig.5. 5) Why would introducing ordinal feature alignment only gives a margin improvement? How is ”m“ set in Equation 5? 6) In the second paragraph of the introductory section, only two articles are used to conclude the limitations, which we think is not very convincing, whether there are other applications of deep learning for ageing fundus pictures, I suggest adding more references. 7) Please keep the” (yj) ̂≠ (yj^,) ̂ “in formula 5 the same as in Fig.2. 8) It is proposed to add the abbreviation “CA” after the class attention in 2.1 to better correspond to the overall framework in Fig.1.
Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making

5
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The paper presents a well-designed and well-executed study on retinal age estimation using deep learning. The proposed method addresses the limitations of previous studies and achieves lower age prediction errors on multiple cohorts. The paper could provide more details on the dataset and implementation and more insights into the biological
Reviewer confidence

Confident but not absolutely certain
[Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

N/A
[Post rebuttal] Please justify your decision

N/A

Review #1

Please describe the contribution of the paper

The authors proposed a progressive label distribution learning that contains a two-stage regression head and a temporal branch to model sequential fundus images and their temporal evolution. The authors used a large dataset with temporal fundus images (130K) and they claim to improve the retinal age estimation on fundus images from multiple cohorts.
Please list the main strengths of the paper; you should write about 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 experimental setup, the large dataset used in this paper, and the ablation study.
Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

I think that the main weakness of this paper is the comparison with other methods, the authors must include other state-of-the-art methods!
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance

The authors should release the code to ensure the reproducibility of the paper
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html

The paper is well written, and the experimental setup, the large dataset used in this paper, and the ablation study are reported in detail but the comparison with other methods reported in the literature is the main weakness.
Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making

5
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The reproducibility of the paper and the comparison with other methods should be reported in detail.
Reviewer confidence

Very confident
[Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

N/A
[Post rebuttal] Please justify your decision

N/A

Primary Meta-Review

Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. 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 progressive label distribution learning to estimate the retinal age with temporal fundus images. Based on the three positive reviews, I recommend this paper as accept. However, there are some concerns arised by the reviewers, such as missed details, unclear difinination, typos. Please revise those in the publish version.

Author Feedback

N/A

back to top

Retinal Age Estimation with Temporal Fundus Images Enhanced Progressive Label Distribution Learning