List of Papers Browse by Subject Areas Author List
Abstract
Medical imaging provides a wealth of information about a patient’s physical condition, and imaging-derived phenotypes (IDPs) extracted from medical images have applications in various biomedical tasks such as disease predic-tion and phenotype association studies. For disease prediction tasks, the col-lection of multimodal imaging data and the conduct of long-term follow-ups are crucial; however, the low incidence rates of certain diseases make it chal-lenging to acquire large-scale cohort data. On the other hand, cohorts that contain genomics and blood-based biomarkers are relatively extensive. Against this backdrop, large-scale cohort data from the UK Biobank (UKB) were leveraged to construct prediction models for 260 IDPs extracted from common brain MRI and cardiac MRI using machine learning methods com-bined with genomics and basic blood characteristics. We applied these mod-els to impute IDPs in cohorts missing imaging data and utilized the imputed IDPs for IDP-disease association studies and disease prediction. Association study results demonstrate that the imputed IDPs can reveal numerous IDP-disease associations. Furthermore, the disease prediction models developed using imputed IDPs demonstrated significantly superior performance across 184 common diseases, as evidenced by higher overall AUC values when compared to models utilizing real IDPs (Wilcoxon signed-rank test, p < 0.001). These results clearly highlight the significant application value of our IDPs prediction models in the context of disease discovery.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2143_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/BioLemon/MultiModal_IDPs_Prediction
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaHao_Multimodal_MICCAI2025,
author = { Zhang, Haoyang and Li, Yan and Liu, Junhong and Lan, Lizhen and Wang, Zian and Sun, Longyu and Lv, Yuntong and Yang, Shengxiao and Li, Qing and Sun, Mengting and Zhang, Yajing and Chen, Binghua and Zhou, Xionghui and Wu, Lianming and Wang, Chengyan},
title = { { Multimodal Imputation of Imaging-derived Phenotypes from Genomic and Blood-based Biomarkers Enhances Common Disease Discovery } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15967},
month = {September},
page = {353 -- 362}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose a method for imputing missing IDPs in the scenario where there are subjects missing all IDPs rather than just a few. A predictive model is trained on data from the UKB and then used to predict the missing IDPS
- 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.
It is an interesting application which would be very useful for many studies The paper figures are good and explain the contributions well The paper utilises good statistics to demonstrate the application
- 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.
The method is only explored for a single dataset - it is trained on UKB and applied to UKB subjects missing IDPS. This is a very limited scenario and it would be much more useful to train on UKB and apply to other datasets where the IDPS are not available. There would however likely be issues due to domain shift so this would need to be explored. However, I think this is essential for the utility of the paper.
The proposed method is very simple, and an exploration of different methods to do the prediction would be important. It isnt obvious that a linear regression model would be the best choice
Quite a few sentences are a bit confused and hard to follow - some examples below but there are others:
-
On the other hand, cohorts that contain ge-nomics and blood-based biomarkers are relatively extensive. - I dont think you mean extensive
-
This arguement isnt very clear: for disease prediction tasks, the collection of multimodal imaging data and long-term follow-up are essential, but are often limited by the low incidence rates of certain diseases.
How were the IDPs selected? This needs explaining
-
- 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.
- 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/2025/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 question being explored is of interest but the work as presented is too limited of scope to be of interest to the MICCAI community
- 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.
The authors have addressed the most of the points raised but my major limitations have not been sufficiently addressed for acceptance. In particular, the limitation of the same training and testing set remains and I feel majorly reduces the utility of the work.
Review #2
- Please describe the contribution of the paper
The paper presents a novel two-stage (two-stack) machine learning framework for imputing imaging-derived phenotypes (IDPs) using genomic and clinical features in the UK Biobank dataset. In the first stage, the model predicts IDPs from non-imaging data (genomics and clinical variables). In the second stage, the predicted (imputed) IDPs are used as features in downstream tasks, such as disease classification and association analysis. This approach aims to enable the use of valuable imaging-derived traits in large-scale studies where imaging data may be missing or unavailable, thereby expanding the utility of imaging-based biomarkers in biobank-scale datasets.
- 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 paper introduces a two-stack model that separates the imputation of imaging-derived phenotypes (IDPs) from their application in downstream clinical tasks.
- 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.
In this manuscript the authors proposed an approach to impute imaging-derived phenotypes (IDPs) from genomic and clinical features in the UK Biobank dataset. The overall idea is interesting; however, there are some comments and questions that need to be addressed:
- The authors used a two-stack model: the first stack predicts IDPs from genomic/clinical data, and the second stack uses the predicted IDPs for downstream tasks. It appears that in the second stack, the authors used 30% of the real IDP population as a test set for both real and imputed IDPs. However, it is unclear what exactly was done during the training phase of the first stack. Did they use the entire real IDP dataset for training the first stack? If so, this would introduce data leakage, as part of this “seen” data would later be used as an “unseen” test set in the second stack. Please provide additional details and clarify the training and testing splits used for both stacks.
- Section 3.1, The results of the first stack, which predicts imputed IDPs, are not particularly promising. Most of the reported R² and correlation coefficients (r values) are low and not reliable.
- In Section 3.2, the authors evaluated the association between the imputed IDPs and disease. However, the setup used to validate the imputed IDPs raises questions. The authors reported correlations between disease-associated real IDPs and imputed IDPs in two different sub-cohorts of the UKbiobank. However both sub-cohorts were derived from the same cohort (UK Biobank), demonstrating correlation between these sub-cohorts does not provide strong evidence for the utility of the imputed IDPs. To better demonstrate the effectiveness of their model, the authors should use a truly unseen test set—e.g., 30% of the real IDP data not used during training—and then evaluate the correlation between the imputed and real IDPs within this held-out subset.
- In Section 3.3, please provide details on how the 184 disease categories were reduced to 13 classes. Were ICD-10 codes used for this categorization?
- Also in Section 3.3, and in line with the earlier comments: the authors should clearly explain how data splitting was handled during the first stack model training. If the test data used in this section was also used during the training of the first stack, this would again result in data leakage. Please clarify and provide a detailed description of the data partitioning strategy.
- Please add your UKbiobank IRB (application number) statement in manuscript.
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
Authors need to add their UKbiobank IRB (application number) statement in manuscript.
- 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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method is conceptually interesting and clinically relevant, offering a scalable solution where imaging data is limited. However, the paper suffers from methodological concerns, particularly the lack of clarity around data splitting and potential data leakage between training and evaluation stages. Additionally, the imputation performance is relatively weak, and the evaluation design does not fully demonstrate generalizability. Despite these limitations, the overall idea is good.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The paper describes an experiment where the authors try to predict 260 image derived phenotypes (IDPs) from blood count data (BC), blood biochemistry data and SNPs. These predicted IDPs are used for other downstream tasks, such as disease prediction.
- 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 is generally understandable 2) The summary figure is good 3) The authors work on an extensive real dataset 4) The authors state their main limitations of the interpretations in the discussion paragraph
- 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 main weakness is that the scope is simply too wide. It would be more valuable to the reader to gain more insights into which particular IDPs are easier to predict, and why it is harder to predict some. The authors don’t talk enough about particular IDPs. The ones that are easier to predict, do they have more SNPs associated? Is it possible to get more insights on why some are harder to predict, e.g., are these IDPs that correspond to small brain regions that are harder to quantify? It would be good to also have the heritability of the IDPs compared how easy it is to predict them based on BC and BBC data, this would give some insight into which have a stronger genetic components and which don’t. 2) It would have been simpler to use polygenic risk scores instead of training the SNP data from scratch. 3) Some predicted IDPs have high variance explained and the reader needs a better summary on whether that is stemming from a single source, or whether there are independent contributions from all the three sources (BC, BBC and SNPs). 4) Some SNPs are super predictive, e.g., the e4 allele in the APOE gene and the MHC region. These should ideally be omitted in an ablation study to quantify how much they contribute to the predictions. 5) There is no mention of code and particular data used. It would be impossible to replicate. What software is used for logistic regression, which covariates are used… etc.
- 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.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors propose an attempt to solve an interesting problem. Imaging data is expensive to acquire and it is interesting to gain a better understanding of how other standard measured values can predict them.
- 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.
The authors provide satisfactory answers to the questions and issues raised by the reviewers and the problem they are trying to solve is interesting. The author’s contribution is relevant for the field of imaging and genetics, since good imputations of IDPs could be relevant for large datasets with limited imaging data.
Author Feedback
Thank you to all reviewers and meta-reviewers for their careful review of our paper. Below, we will address the reviewers’ concerns point by point. 1.Datasets split (R1) To clarify, the two-layer stacked model in our paper is only used for building the IDP prediction models and didn’t participate in downstream tasks. The two-layer model designed specifically for building single-modality BR models and fusion models. We split the UKB dataset into two independent subsets: one containing IDPs (28,615 subjects) and the other without IDPs (305,055 subjects). The former was used for constructing the IDP prediction models, while the latter was employed to evaluate the application value of predicted IDPs data in downstream tasks. Thus, no information leakage occurred. 2.Usage of simple structure and achieve with relatively low R2/r (R1, R3) (1)Predicting IDPs using SNPs and blood characteristics is a challenging task. Therefore, the current R2/r in this field are not low, and such performance already enables predicted IDPs to perform well in downstream tasks. (2)Our downstream tasks focus on a disease risk prediction. IDPs exhibit significant differences between healthy individuals and patients. Even if the R²/r values are not very high, they can still effectively distinguish between healthy individuals and patients, as shown in our results. (3)Due to the need to construct models for a large number of IDPs, we require simple and effective model architectures, which is why we adopted the current BR-based linear model. In the future, we will compare the performance of multiple models. 3.Why not PRS (R2) As stated in reference [11] of our paper, constructing PRS using weights trained via BR outperforms the traditional approach using GWAS β-values due to its supervised nature. Therefore, we used BR to construct the prediction model for the SNP modality, merely without organizing it into a genetic score file. 4.Correlation of IDPs and disease(R1) We agree that it is necessary to reserve some data with IDP gold standard for model validation. However, for most diseases, 30% of samples with real IDPs contain too few patients (e.g., a 0.1% incidence disease yields ~9 cases from 28,615 × 0.3 × 0.001), which would make statistical results unreliable. In the future, we will conduct the validation experiments you mentioned on some high-incidence diseases. 5.Disease categorization(R1) Diseases are grouped by standard ICD-10 chapters. 6.Gain insights into particular IDPs, summary on variance explained and ablation study(R2) Thank you for your valuable comments! Due to limited space, we were unable to include experiments you have mentioned in our paper, but we will elaborate on them in the Limitations section. We believe the current paper already fully demonstrates the feasibility of predicting IDPs using SNPs and blood characteristics, as well as their application value in downstream tasks. 7.Multi-dataset validation(R3) The UK Biobank is a multi-center large-scale biobank containing hundreds of thousands of samples. Although we did not partition the dataset by center, we believe the dataset currently used already has a high level of diversity. 8.Interest to the MICCAI community(R3) Our study holds application value for two types of datasets: those containing SNP and blood characteristic data but lacking IDPs, and those with only IDPs but insufficient sample size. Researchers engaged in image analysis can utilize the disease prediction model parameters constructed based on predicted IDPs in Section 3.3. Notably, these parameters exhibit better disease prediction efficacy than those derived from the UKB cohort with real IDPs. 9.Reproduce information(R1, R2, R3) We will make the code and data publicly available on GitHub after paper publication. 10.Selection of IDPs(R3) We selected IDPs of two organs with high research attention, rich phenotypic traits, and abundant GWAS resources.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
Please take account of the reviewers comments. Please not that adding the UKbiobank IRB (application number) is not required for this anomymized version intended for review.
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A