Abstract

Opportunistic osteoporosis screening using routine CT scans enables bone mineral density (BMD) assessment without additional imaging modalities. However, conversion from Hounsfield Units (HU) to physical density requires calibration, and scanner-dependent variations limit generalizability across clinical centers. We propose a supervised 3D image to-image translation framework that learns to synthesize voxel-wise Quantitative Computed Tomography (QCT) density maps of the proximal femur directly from routine CT. For clinical evaluation, volumetric BMD (vBMD) is computed as the mean calibrated density within the proximal femur mask. Agreement is assessed between predicted vBMD and phantom-calibrated reference vBMD using mean absolute error (MAE), Pearson correlation coefficient (PCC), and intra-class correlation coefficient (ICC). We evaluate the method on a large multi-center dataset of 2157 patients from 9 centers, using leave-one-center-out cross-validation. Centers with a similar calibration factors distribution are considered in distribution (ID), while centers with distinct calibration factors are considered out-of-distribution (OOD), representing a more challenging scenario for model generalization. Performance is compared against conventional calibration baselines. Simply applying the mean calibration factors of the training dataset achieve high agreement on ID data (ICC = 0.981, MAE = 10.0 ± 8.6), but their reliability deteriorates on OOD data (ICC =0.620, MAE=40.8±7.4). In contrast, our image-to-image translation model maintains robust agreement on OOD data (ICC = 0.877, MAE=19.1 ± 10.5), demonstrating resilience to center-specific calibration shifts and improved generalizability across clinical settings.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5051_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Jobs27/QCT-translation

Link to the Dataset(s)

N/A

BibTex

@InProceedings{SchJob_Learning_MICCAI2026,
        author = { Schöller, Job H. J. AND Gu, Yi AND Uemura, Keisuke AND Takao, Masaki AND Okada, Seiji AND Sugano, Nobuhiko AND Sato, Yoshinobu AND Otake, Yoshito},
        title = { { Learning Quantitative Calibration of Computed Tomography for Bone Mineral Density Estimation Using Multi-Center Data } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16881},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a quantitative framework for estimating bone mineral density (BMD) using Quantitative Computed Tomography (qCT). A phantom-calibrated reference is employed to evaluate the accuracy of the predicted volumetric BMD (vBMD), ensuring that the measurements are physically meaningful rather than purely intensity-based.

    The method formulates BMD estimation as an image-to-image translation task, where standard CT scans of the proximal femur are processed after segmentation to generate QCT-equivalent outputs. This implicitly attempts to recover calibrated density information from non-calibrated or heterogeneous CT inputs.

    Evaluation is conducted on a multi-center dataset, assessing both in-distribution (ID) and out-of-distribution (OOD) generalization. The results indicate that the proposed approach achieves improved performance compared to state-of-the-art methods across both settings, suggesting robustness to scanner variability and acquisition differences.

  • 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 proposed method leverages QCT with ground-truth volumetric bone mineral density (vBMD) obtained via phantom-based calibration, ensuring physically meaningful density measurements. The approach is evaluated on a large multi-center dataset, employing a leave-one-center-out validation strategy to rigorously assess generalization across different acquisition settings. The method is compared against several calibration-based baseline approaches, and results demonstrate improved performance, particularly in out-of-distribution (OOD) centers. This indicates that the proposed image-to-image translation framework is effective in mitigating inter-scanner variability and recovering QCT-equivalent BMD from heterogeneous CT inputs.

  • 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 image to image translation approach instead of physics based calibration as qCT relies on scanner specific calibration as learning approach approximates BMD but don’t always guarantee clinical correctness The pipeline is based on proximal femur segmentation but getting accurate segmentation masks is not given in the paper and this also could limit applicability to other bone anatomical regions as error in segmentation could propagate to BMD estimation as well. Also comparions with other learning based BMD estimation methods should have been made and also extreme contrast shifts on CT as each center would be using same contrast level CT which again could impact the learning based method performance.

  • 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?

    From an experimental standpoint, the comparisons are limited to relatively simple baselines such as mean and regression-based calibration. The absence of benchmarking against other learning-based or domain adaptation methods weakens the claims of improved generalizability. Additionally, the use of a single out-of-distribution center does not adequately capture the full spectrum of inter-scanner and inter-center variability encountered in clinical practice. Overall, despite leveraging a large and diverse dataset, these concerns around limited novelty, dependency on calibration during training, and insufficient comparative evaluation make the work fall short of the bar for acceptance, supporting a weak reject decision.

  • 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 #2

  • Please describe the contribution of the paper

    The paper proposes a method to estimate volumetric bone mineral density (vBMD) from routine CT scans by learning to generate quantitative CT images using 3D image translation. Computation of vBMD from opportunistically available CT scans would improve accessibility for especially osteoporosis care.

    The authors test their model on in-distribution and out-of-distribution data from multiple clinical centres, one of which differs significantly from the rest. They demonstrate that their proposed approach performs competitively with baselines on the similar clinics, but outperforms the baselines on the out-of-distribution clinic.

  • 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 authors propose a relatively simple approach based on image translation, converting routine CTs to QCTs, extracting the vBMD post-hoc. The baseline comparisons are exhaustive and well-described. They cover the most intuitive alternative ways to compute and regress the vBMD, showing that the baselines mostly mimic the training data regression parameters. I fail to see that more complicated deep learning approaches would perform much differently.

    Their main strength is the insight that some clinics/machines may exhibit vastly different calibration parameters and the demonstration of their method’s out-of-distribution generalization. The method of evaluation is very relevant, showing the reduced loss of generalizability on patients from clinics with substantially different parameters compared to other methods.

    Likewise, the paper contains an ablation of some of the most relevant factors that likely impact model performance.

    The quality of writing is overall good and the thesis is sufficiently motivated.

  • 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 paper may suffer slightly from “carrying coals to Newcastle”, doing something that may be complicated and ultimately unnecessary. I would have liked to see discussion of why their method works - they are essentially estimating a linear relationship, which apparently may most easily be done through deep learning-based image-to-image translation…

    The paper lacks more extensive discussion on the correctness and validity of the ground truth vBMD extracted in this fashion from CTs, and the ultimate utility of their method for using the evaluated vBMDs in a downstream task, e.g. fracture survival analysis or osteoporosis risk. Such analysis would have further increased the impact.

    As for the method, image-to-image translation is not new, and similar efforts exists to convert between modalities, e.g. CT, MRI, or the similar task of radiotherapy dose estimation from MRI or other complementary modalities. The authors should highlight similarities and deviations from this kind of previous work. Finally, I think that it would still have been relevant to withhold data from the in-distribution to see how well it generalizes from training to test on that set.

    Results-wise, the performance of their method is solid, but as shown in Fig. 3, there is a systematic underestimation of the ground truth vBMD across all methods, and it would have benefited the paper to address this more explicitly, and to propose overall fixes. Additionally, it is right now not clear exactly how the mean+-std was computed, and the deviations are comparatively large.

    I would also suggest restructuring the Experiments section with e.g. \paragraph to emphasize the different baseline models.

    I also suggest displaying results in terms of MAPE (mean absolue percentage error) which is often useful in regression tasks.

  • 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.

    (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 paper is in general solid and the experiments sufficient for publication, but the discussion of their method and results is lacking.

    I would like to see more references to similar, relevant methods of image-to-image translation, further discussion on the validity of the vBMD results and their utility (e.g. using relevant references), as well as expanded analysis of the vBMD under-estimation and potential lines of future work to adress this. Lastly, the computation of variability presented in the results should be clarified.

  • 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 have sufficiently addressed my concerns.



Review #3

  • Please describe the contribution of the paper

    The main contribution of the paper is that it demonstrates that QCT having vBMD information can be derived with a simple U-Net neural network algorithm using a phantom-less calibration approach, at least within their dataset covering 9 sites. The second major contribution of the paper is the collection of the QCT dataset from nine sites for U-Net training. Phantom-less calibration approaches where volumetric BMD is derived for CT are needed to produce quantitative CT without a physical calibration phantom being present in the CT scan. vBMD information for CT is needed in several tasks, especially in CT-based finite element modeling for calculating the fracture strength of the bone and in opportunistic screening for osteoporosis. The topic of the paper is highly relevant since the opportunistic screening of decreased bone mineral density and elevated fracture risk is a highly potential approach to tackle the challenge of osteoporosis underdiagnosis. Patients directed to radiological examination are at higher risk for fracture than the normal population, so it makes sense to aim screening at them.

  • 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 study addresses a significant research question that is not yet fully solved in the previous literature. The study shows that a straightforward and simple U-Net algorithm should be enough to calibrate vBMD in CT images. Nine sites are a strong indicator of the method’s generalizability, even though the study had limitations addressed below. The strength of the proposed method is in estimating out-of-distribution data, where it outperforms baseline methods such as mean calibration from training data. I especially liked figure 1, which gave a good overview with one seek the range of vBMD and the calibration coefficients variation within and between sites. It also illustrates the FOW where calibration was done. Also, the schematic overview in Fig. 2 was clear.

  • 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 description of prior calibration techniques and the research gap is limited. Even though this is a conference paper, it would have benefited from a concise summary of the earlier methods and their limitations. Approaches often segment different tissue types to calibrate the mapping from HU to vBMD (e.g. reviewed in https://doi.org/10.1002/jbmr.4917). Using U-NET directly is in some sense straightforward, whereas other approaches have high transparency in deriving the vBMD calibration function. Now these issues are not addressed. Authors did not explain why they cropped the image strictly to femur. Previous literature (e.g. https://doi.org/10.1002/jbmr.4917) has shown that information from different tissue types, such as fat and blood in veins and air around the body can be used in calibration. Intuitively, cropping only to femur may limit the method’s capability to calibrate since it can less use information from different tissues. It also remains to be seen how the calibration functions on other bones than bone, since they are not now covered. Authors did not give information about what CT scanner manufacturers and models were covered in their dataset and whether different sites used CTs from different manufacturers, or different imaging voltages, reconstruction kernels or resolutions. Therefore, it remains unclear how well the spectrum of devices and settings from different manufacturers was covered in the study. It remained unclear why Center9 differed so strongly from other sites. The study would have benefited if the authors had discussed potential reasons for the out of distribution results. Ablation study: It is debatable if one can state that Exp. D did not yield better results than Exp. A but still say Exp. C provided a small improvement, when MAE and ICC of C and D are closer to each other than those of A and D. The study did not include algorithm development, not sure if this is a limitation, though, but the study focused on showing that a simple U-Net should work for vBMD calibration.

  • 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.

    (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 study did not include algorithm development, but the study focused on showing that a simple U-Net should work for vBMD calibration. I decided to give weak accept since the study motivates to consider simple direct U-net based approach when one searches for a way for phantom-less vBMD calibration of CT images. Knowledge of the covered CT manufacturers, models and protocol settings would strengthen the study.

  • 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 rebuttal adequately addressed my concerns.

    The study addresses an important and still partly open research problem in the literature. The results suggest that a relatively straightforward U-Net-based model may be sufficient for vBMD calibration in CT images. The inclusion of data from nine sites provides encouraging evidence for the potential generalizability of the method. These results have clear practical applications, particularly in fracture risk estimation based on finite element modeling.

    The authors have committed to providing more detailed information on the CT scanner manufacturers and models included in the dataset, as well as their potential impact on calibration accuracy. This is particularly important for interpreting the notably different behavior observed at Center 9, and may help readers to evaluate method’s generalizability.

    While some uncertainty remains regarding cross-scanner variability, generalizability to other anatomies and robustness in case of imprecision in the initial cropping of the anatomy of interest, the work provides a compelling argument for considering simple, direct U-Net–based approaches for phantom-less vBMD calibration.

    Based on these considerations, I recommend acceptance.



Author Feedback

We thank the reviewers for their constructive feedback. Below, we address the main concerns.

Q1 (R1, R2, R3) Comparison with existing methods We agree that the relationship between our approach and prior methods for BMD estimation should be discussed more comprehensively. Our approach differs fundamentally by formulating the problem as a spatial I2I translation task that predicts voxel-wise QCT density maps rather than regressing BMD [9] or estimating calibration factors [1]. We would like to clarify that the use of a simple U-Net and standard loss functions was intentional, serving as a proof-of-concept to evaluate proposed formulation itself. We will clarify this point more explicitly in the revised paper. [1] S. Bartenschlager et al.: Opportunistic Screening With CT: Comparison of Phantomless BMD Calibration Methods. JBMR. (2023)

Q2 (R1, R2): Conceptual motivation and theoretical justification Our central hypothesis is that the proposed I2I translation framework learns a scanner normalized representation of physical bone density, whereas direct regression only learns a global statistical shortcut. The task is fundamentally spatial and benefits from dense voxel-wise supervision, which provides vastly richer learning signals than direct regression based on single scalar values. We will clarify this point more clearly in the revised paper.

Q3 (R1, R2): Validity of ground truth We apologize for the insufficient information regarding the ground-truth (GT) BMD generation. The GT-BMD was generated following the existing pipeline consisting of CT bone segmentation, phantom segmentation, and BMD calculation [2]. We assume the proximal femur (PF) segmentation is readily available before applying our method. We will clarify the details of GT-BMD generation in the revised paper. [2] K. Uemura et al.: Development of an open-source measurement system to assess the areal bone mineral density of the proximal femur from clinical CT images. Arch Osteoporos. (2022)

Q4: (R1, R2): In-distribution (ID) and out-of-distribution (OOD) setting We apologize for the lack of clarity regarding the experimental protocol. All nine centers were evaluated using a leave-one-center-out cross-validation strategy. After completing all folds, the test results were grouped into ID and OOD categories based on the calibration-factor distributions. Thus, the results aggregate outcomes across multiple cross-validation experiments. The revised paper will provide a clearer description of this setup.

Q5 (R2): systematic underestimation of BMD We agree that the proposed framework still exhibits systematic underestimation of BMD for the OOD dataset, and we will explicitly acknowledge this as a limitation in the revised paper. At the same time, our results demonstrate that BMD regression or calibration estimation approaches are substantially less robust, whereas the proposed method maintains high agreement in the OOD setting. We will further expand this discussion in the revised paper.

Q6 (R3): CT acquisition and scanner variability We agree that additional details regarding CT acquisition settings would strengthen the manuscript. The dataset includes scanners from GE and Canon, and Center 9 used a lower-dose acquisition protocol, which may contribute to the observed OOD calibration shift. We will provide additional scanner and acquisition details in the revised paper.

Q7 (R3): Use of the proximal femur cropped region Although the input volumes are cropped around the PF region, surrounding anatomical structures and soft tissue remain partially visible and may still provide contextual information to the prediction. The cropping strategy was intended to focus the model on the clinically relevant region while retaining limited contextual cues. We will clarify this design choice more explicitly in the revised paper.

Other comments: The revised paper will add MAPE as an additional metric (R2). Our code will be made available upon acceptance (R1, R2, R3).




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 receives mixed reviews. Reviewers raise major concerns of this work, such as limited novelty, dependency on calibration during training, and insufficient comparative evaluation. Rebuttal is invited.

  • 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 received mixed reviews. Based on all reviewers’ comments and author response, I agree that this paper does not provide technical novelty, however, this work provides a compelling argument for considering simple, direct U-Net–based approaches for phantom-less vBMD calibration. It is also validated using data from 9 centers. Hence, I lean to acceptance of this work.



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.

    The novelty of the method is very low (it just uses a U-Net). However, the application is relevant, and the results show that a simple method can yield results that can be used in clinical studies. The tests on 9 different sites and machines strengthen the results.



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.

    Important notice: despite a reminder, reviewer 1 failed to submit a post-rebuttal evaluation.

    The paper presents a methodology which is very nicely evaluated in a multicenter setup to provide a valuable clinical solution. The simplicity of the approach is definitely appealing. Also, assuming the authors provide the promised information and implement the adjustments on the camera-ready paper, the paper would be a very nice paper for the main conference.



back to top