A Functionally Guided U-Net for Chronic Kidney Disease Assessment: Joint Structural Segmentation and eGFR Prediction with a Structure–Function Consistency Loss
Round 1
Reviewer 1 Report
Comments and Suggestions for Authors- It is commendable to present new predictive models based on machine learning including diagnostic procedures.
- The agreement rate of this research is 14% confirmed by iThenticate which is more than commendable!
The disadvantages are that the introduction is too broad. It is necessary to significantly reduce the number of words by as much as 60%. Images are not needed in the introduction. It is necessary to change the colors in the images Figure 2.
The research is excellently conducted. I have no complaints.
Author Response
Reviewer 1
Comment:
- It is commendable to present new predictive models based on machine learning including diagnostic procedures. - The agreement rate of this research is 14% confirmed by iThenticate which is more than commendable! The disadvantages are that the introduction is too broad. It is necessary to significantly reduce the number of words by as much as 60%. Images are not needed in the introduction. It is necessary to change the colors in the images Figure 2. The research is excellently conducted. I have no complaints.
Response:
We sincerely thank the reviewer for the positive and encouraging evaluation of our work, as well as for recognizing both the originality of the proposed machine learning–based diagnostic framework and the low similarity index confirmed by iThenticate. We greatly appreciate the reviewer’s constructive comments, which have helped us improve the clarity and presentation of the manuscript.
In response to the reviewer’s suggestions, we have significantly revised the Introduction section, reducing its length by approximately 60% while preserving the core motivation, relevant literature, and the positioning of the proposed method. This revision ensures a more focused and concise introduction that directly highlights the research gap and contributions.
In addition, figure 1 has been removed, in accordance with the reviewer’s recommendation. These changes streamline the introductory section and improve the overall manuscript structure.
Furthermore, we have updated the color scheme of Figure 2, adopting a more neutral and publication-appropriate palette to enhance visual clarity and consistency with academic standards.
Once again, we sincerely thank the reviewer for the valuable feedback and for the positive assessment of the research quality. All suggested revisions have been fully addressed in the revised manuscript.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe article "A Functionally-Guided U-Net for Chronic Kidney Disease Assessment: Joint Structural Segmentation and eGFR Prediction with a Structure-Function Consistency Loss" presents a novel multitask deep learning framework that integrates kidney compartment segmentation with eGFR prediction, using a structure–function consistency loss. The approach is clinically motivated and contributes to the growing field of renal disease assessment.
Key suggestions for improvement include the following.
The eGFR values are derived from CKD stage midpoints rather than laboratory-measured serum creatinine and cystatin C. The authors could consider stating that eGFR labels are estimated from stages, not measured, and discuss the implications in the Limitations section.
If possible, incorporate real laboratory eGFR values or use established formulas (CKD-EPI, MDRD) for a subset of subjects to validate the regression output.
The authors should also acknowledge the small MRI dataset sample size which may not capture the full spectrum of CKD etiologies and morphological diversity. Also, there might be a potential selection bias, considering the fact that the CT dataset includes non-CKD pathologies that are not representative of typical CKD progression, like cysts or stones, which should also me acknowledged as a limitation.
eGFR is severely influenced by age, sex, race, and comorbidities. Since the model in this study does not incorporate these factors, the authors are encouraged to mention this limitation as well
Also, the authors should insert a discussion on potential for longitudinal extension (e.g., predicting eGFR slope), considering the fact that CKD is a progressive disease.
In the Discussion section, the authors should consider proposing concrete clinical use cases, address integration into clinical workflows, and suggest examples where structural biomarkers can clearly explain functional decline.
Other than the mentioned, no further adjustments are necessary.
Author Response
Reviewer 2
Comment 1:
The eGFR values are derived from CKD stage midpoints rather than laboratory-measured serum creatinine and cystatin C. The authors could consider stating that eGFR labels are estimated from stages, not measured, and discuss the implications in the Limitations section.
Response:
We thank the reviewer for this insightful and constructive comment. We agree that clarifying the nature of the eGFR labels is important for accurate interpretation of the results.
In response, we have explicitly stated in the manuscript that eGFR values are estimated from CKD stage midpoints rather than directly measured laboratory biomarkers such as serum creatinine or cystatin C. Additionally, we have expanded the Limitations section to discuss the implications of using stage-derived eGFR values, including the lack of intra-stage variability and the interpretation of predictions as relative functional indicators rather than exact clinical measurements. These clarifications improve transparency and better contextualize the scope and applicability of the proposed method.
We appreciate the reviewer’s suggestion, which has strengthened the methodological clarity of the manuscript.
Comment 2:
If possible, incorporate real laboratory eGFR values or use established formulas (CKD-EPI, MDRD) for a subset of subjects to validate the regression output.
Response:
We thank the reviewer for this valuable suggestion regarding the use of laboratory-derived eGFR values or established clinical equations for validating the regression output.
We agree that incorporating real laboratory measurements (e.g., serum creatinine or cystatin C) and computing eGFR using standard formulas such as CKD-EPI or MDRD would provide the strongest form of functional validation. However, in the datasets used in this study, subject-specific laboratory measurements are not available, which makes direct computation of biochemistry-based eGFR infeasible. This limitation is common in publicly available renal imaging datasets, where functional annotations are often provided only at the CKD stage level rather than as raw laboratory values.
To address this constraint in a principled and clinically grounded manner, we followed the same estimation strategy adopted in [31], where continuous eGFR values are derived from clinically accepted CKD-stage definitions and validated diffusion-based regression modeling. This approach enables consistent functional labeling while preserving physiological ordering across disease severity levels. We have now explicitly clarified this choice in the manuscript and discussed its implications in the Limitations section, emphasizing that the predicted eGFR values represent stage-consistent functional estimates rather than laboratory-measured filtration rates.
Comment 3:
The authors should also acknowledge the small MRI dataset sample size which may not capture the full spectrum of CKD etiologies and morphological diversity. Also, there might be a potential selection bias, considering the fact that the CT dataset includes non-CKD pathologies that are not representative of typical CKD progression, like cysts or stones, which should also me acknowledged as a limitation.
Response:
We thank the reviewer for this insightful and important comment regarding dataset size and potential selection bias.
We fully agree that the limited MRI sample size may not encompass the full diversity of CKD etiologies, morphological patterns, and disease trajectories encountered in broader clinical populations. In response, we have explicitly acknowledged this limitation in the revised manuscript and clarified its potential impact on model generalizability.
We also agree that the CT dataset includes non-CKD renal pathologies (e.g., cysts, stones, tumors) that are not representative of typical CKD progression. As clarified in the manuscript, this dataset was used exclusively for assessing structural robustness and cross-modality generalization, not for CKD-specific functional prediction. Nevertheless, we now explicitly acknowledge the possibility of selection bias arising from this heterogeneity and discuss its implications in the Limitations section.
Comment 4:
eGFR is severely influenced by age, sex, race, and comorbidities. Since the model in this study does not incorporate these factors, the authors are encouraged to mention this limitation as well.
Response:
We thank the reviewer for raising this important point regarding the influence of demographic and clinical factors on eGFR estimation.
We would like to clarify that age and sex were explicitly incorporated into the functional modeling pipeline, as provided in the dataset metadata and illustrated in the attached dataset description (see the demographic attribute summary). These variables were included in the regression formulation used to derive and validate eGFR estimates, consistent with established clinical practice. This integration is explicitly described in the manuscript in the paragraph below Figure 9 and immediately above Equation (33), where the multivariate formulation and contributing variables are detailed.
Regarding race and comorbidities, we acknowledge that these variables were not available in the utilized public datasets and therefore could not be incorporated into the current model. This constraint is now clearly stated as a limitation in the revised manuscript. We emphasize that this limitation reflects dataset availability rather than methodological omission, and we note that future work will prioritize cohorts containing richer demographic and clinical metadata to further enhance physiological fidelity and fairness.
We appreciate the reviewer’s comment, which allowed us to clarify this aspect and improve the transparency of the manuscript.
Comments 5 and 6:
- Also, the authors should insert a discussion on potential for longitudinal extension (e.g., predicting eGFR slope), considering the fact that CKD is a progressive disease.
- In the Discussion section, the authors should consider proposing concrete clinical use cases, address integration into clinical workflows, and suggest examples where structural biomarkers can clearly explain functional decline.
Response:
We thank the reviewer for this constructive and forward-looking comment regarding the longitudinal nature of CKD and the clinical applicability of the proposed framework.
In response, we have added a new dedicated subsection (Section 4.2) to the Discussion that explicitly addresses the potential for longitudinal extension of the proposed FG-CKD-UNet. In this section, we describe how the framework can be naturally extended to model eGFR trajectories and eGFR slope by incorporating sequential imaging data and temporal learning modules, enabling prediction of disease progression rather than single-time-point severity.
Furthermore, we have expanded the Discussion to include concrete clinical use cases and considerations for integration into real-world clinical workflows. Specifically, we discuss how the proposed model can be deployed as a post-processing tool within standard radiology pipelines, integrated with PACS systems, and used to support joint radiology–nephrology decision-making. We also provide clear examples of how structure-derived biomarkers (e.g., cortical thinning, parenchymal volume loss, cortex–medulla imbalance) can offer transparent and clinically meaningful explanations for functional decline, enhancing interpretability and clinical trust.
These additions strengthen the translational relevance of the manuscript and directly address the reviewer’s suggestion. We appreciate this comment, which helped us improve the clinical framing and future impact of the proposed work.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsPlease refer to the attached file for details.
Comments for author File: Comments.pdf
Author Response
Reviewer 3
Comment 1:
The manuscript predicts eGFR as a continuous functional outcome; however, the source and definition of the eGFR labels are not sufficiently clarified. It would be helpful to explicitly state whether these values are derived from clinical laboratory measurements (e.g., serum creatinine–based equations) or inferred from CKD stage annotations. Providing this clarification at the dataset level would improve transparency and strengthen the clinical validity of the study.
Response:
We thank the reviewer for this important comment highlighting the need for clearer clarification regarding the source and definition of the eGFR labels used in this study.
In response, we have explicitly clarified at the dataset and methodology levels that the eGFR values predicted in this work are not derived from direct laboratory measurements (e.g., serum creatinine– or cystatin C–based equations). Instead, continuous eGFR labels were inferred from clinically assigned CKD stage annotations using KDIGO-defined physiological intervals, with midpoint assignment employed to enable regression-based learning. This clarification has been added to the manuscript to improve transparency and ensure accurate clinical interpretation.
Furthermore, we have expanded the Limitations section to discuss the implications of using stage-derived eGFR values, emphasizing that the predicted outputs represent relative functional severity rather than exact biochemical filtration rates. We also outline how future work will validate the regression output using laboratory-confirmed eGFR or established clinical equations (e.g., CKD-EPI or MDRD) when such data are available.
We appreciate the reviewer’s comment, which has helped strengthen the clarity, transparency, and clinical contextualization of the study.
Comment 2:
It is important to emphasize that the morphology-derived eGFR in the proposed framework is not intended to replace or function as a standalone clinical eGFR estimator. Instead, it serves as an auxiliary regularization signal used to enforce physiological consistency between anatomical segmentation outputs and the primary function prediction head during training.
Response:
We thank the reviewer for this insightful comment regarding the role of the morphology-derived eGFR and the potential for misinterpretation.
We would like to clarify that, throughout the manuscript, the primary eGFR prediction is produced exclusively by the functional prediction head, while the morphology-derived eGFR is introduced solely as an auxiliary estimate within the structure–function consistency loss. Its purpose is to enforce physiological alignment between anatomical segmentation outputs and functional predictions, rather than to act as a competing or standalone clinical estimator. This design choice is also reflected in the ablation studies, where the morphology-derived eGFR is treated strictly as a training constraint, not as an independent predictor.
To avoid any possible ambiguity, we have now made this role explicit in the manuscript by adding the following statement above Equation (33):
“It is important to emphasize that the morphology-derived eGFR in the proposed framework is not intended to replace or function as a standalone clinical eGFR estimator. Instead, it serves as an auxiliary regularization signal used to enforce physiological consistency between anatomical segmentation outputs and the primary function prediction head during training.”
This clarification ensures that the intended function of the morphology-derived eGFR is unambiguous and aligned with the overall design of the framework. We appreciate the reviewer’s comment, which helped improve the clarity and interpretability of the proposed method.
Comment 3:
The structure–function consistency loss and multitask objective rely on several weighting parameters (α, β, λ₁–λ₃). While these values appear to be empirically selected, a brief explanation of their tuning strategy (e.g., ablation trends or stability considerations) would enhance reproducibility and methodological rigor
Response:
We thank the reviewer for this constructive comment regarding the tuning of the weighting parameters in the multitask objective.
In response, we have added a brief explanation to the manuscript describing how the parameters , , and – were selected directly above section 3.4. Specifically, we clarify that these weights were empirically tuned based on training stability, convergence behavior, and balanced performance across segmentation accuracy, functional prediction error, and structure–function consistency, rather than arbitrary selection. We also note that ablation trends informed the final values by identifying regimes that either destabilized training or weakened structure–function coupling.
This addition improves methodological transparency and reproducibility, and we appreciate the reviewer’s suggestion, which helped strengthen the rigor of the study.
Comment 4:
Although quantitative performance metrics such as MAE and RMSE are reported, the manuscript would benefit from a short discussion of their clinical relevance. For example, contextualizing the prediction error in relation to CKD stage boundaries or clinically acceptable eGFR variability would provide additional insight into the practical significance of the results.
Response:
We thank the reviewer for this valuable comment regarding the clinical interpretation of the reported regression metrics.
In response, we have added a discussion paragraph that contextualizes the MAE and RMSE values with respect to CKD stage boundaries and clinically acceptable eGFR variability below figure 7 and above table 7. Specifically, we explain how the magnitude of the reported errors compares to inter-stage thresholds and natural physiological fluctuations, and why this level of error is unlikely to result in clinically significant misclassification. This clarification provides additional insight into the practical significance and clinical relevance of the proposed model’s predictive performance.
We appreciate the reviewer’s suggestion, which has strengthened the translational interpretation of the results.
Comment 5:
The manuscript states that the CT dataset was used exclusively for evaluating structural robustness and cross-modality generalization, not for functional prediction. However, this distinction could be more explicitly emphasized in the Methods or Dataset subsection to prevent confusion, particularly given that both MRI and CT images pass through the shared encoder.
Response:
We thank the reviewer for this important clarification request regarding the use of the CT dataset.
In response, we have explicitly emphasized in the Dataset section (3.2) that the CT dataset was employed only for structural segmentation robustness and cross-modality generalization, and that no CT images were used for functional prediction, eGFR regression, or structure–function consistency training. We further clarify that, although MRI and CT images share the same encoder, functional learning and evaluation are conducted exclusively on MRI data with CKD-specific annotations.
This clarification prevents potential confusion and strengthens the transparency of the experimental design. We appreciate the reviewer’s comment, which helped improve methodological clarity.
Comment 6:
Figures 2–4 effectively illustrate the proposed framework but contain a high level of detail. To improve readability, especially in the printed version, the authors may consider moving some algorithmic flow diagrams or detailed subcomponents to an Appendix or Supplementary Materials section.
Response:
We thank the reviewer for this helpful suggestion regarding the presentation and readability of Figures 2–4.
We agree that these figures contain a high level of technical detail, which may affect readability in printed formats. In response, we would like to clarify that the layout and placement of detailed algorithmic flow diagrams and subcomponents will be further refined during the production and proofreading stages, in close cooperation with the journal’s editorial team and the assigned manuscript handler, following acceptance of the manuscript. At that stage, figures can be optimally reorganized, resized, or partially moved to Appendix or Supplementary Materials, in accordance with the journal’s formatting guidelines and editorial recommendations.
We appreciate the reviewer’s comment, which will help improve the final presentation quality of the article.
Author Response File: Author Response.pdf
