List of Papers Browse by Subject Areas Author List
Abstract
3D cell volumes provide richer structural information than 2D images of cells. Therefore, the use of the 3D cell volumes enables stable and reliable cytological diagnosis. Previous work has introduced a Deep Image Prior (DIP)-based method for reconstructing 3D cell volumes from multi-focus image sequences. Although the method does not require a large-scale training dataset, it requires a long time to reconstruct 3D volumes. In this paper, we propose an efficient deep learning-based approach for reconstructing 3D cell volumes from multi-focus image sequences via cross-modality knowledge distillation, termed Deep Cell Volume Reconstruction (DeepCVR). To begin with, using 3D cell volumes, we construct a U-Net-based variational autoencoder, called CellVAE, which learns an optimal latent space for representing the structural information of 3D cell volumes. Next, using the encoder part of the trained CellVAE as a teacher model, we construct a student model, called MF-CellEncoder, which reproduces the teacher’s latent variables from multi-focus images. Finally, DeepCVR is constructed by combining the MF-CellEncoder and CellVAE decoder. DeepCVR reconstructs 3D cell volumes in a single forward pass. This avoids the per-sample optimization required by the DIP-based method. Experimental results show that DeepCVR greatly reduces the reconstruction time while approximating DIP-based reconstructions with comparable quality. Code available at: https://github.com/imamkhairi/DeepCVR.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4794_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/4794_supp.zip
Link to the Code Repository
https://github.com/imamkhairi/DeepCVR
Link to the Dataset(s)
Cervical Cancer Cell Image Database: Multi-focus Cytology Dataset (CCCID): https://zenodo.org/records/18904734
Note: The full clinical dataset used in this study cannot be publicly released due to privacy restrictions; only sample images are publicly available.
BibTex
@InProceedings{LubIma_Fast_MICCAI2026,
author = { Lubis, Imam Khairi AND Morooka, Ken'ichi AND Matsukado, Rin AND Nakamae, Yukako AND Nagahara, Hajime},
title = { { Fast 3D Cell Volume Reconstruction from Multi-focus Images by Cross-modality Knowledge Distillation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors describe 3D volume reconstruction from multifocus images using a teacher model involving a variational autoencoder (VAE) and U-Net-based an encoder model as a student. Combined, they form a 3D volume reconstruction network.
- 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.
Compared to Deep Image Prior, the model provides a significant reduction in reconstruction time - enabling applications outside the scope of the prior art. The selection of models is appropriate for the task, and the methods are used within their scope.
- 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 simulated multifocus images are consistent, but reconstructed volumes seem visually blurred. As the VAE forms a crucial part of the network, additional latent-space statistics could help confirm whether the model uses the latent variables effectively, as posterior collapse cannot be evaluated solely from the images. The stack is padded in z-direction from 11 to 12.A short clarification of how the padded slices are treated downstream would be helpful, as the percentage of added volume is not entirely negligible. While not in the main focus of the paper, short expansion on the cytological categories would be helpful, as they are referred to only as abbreviations. There is unnecessary repetition of dimensions in the beginning of Methodology, which likely could be compacted.
- 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 method shows promise with substantial decreases in computational time. However, due to the nature of generative algorithms, better understanding of how the latent space is handled requires attention.
- Reviewer confidence
Somewhat confident (2)
- [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 answered my concerns related to the latent space. The presented solution provides substantial computational advantages when compared to existing solutions. Although the full justification of architecture is difficult to portray in a 8 page paper, with the clarifications provided by the authors, I see no major issues that would prevent acceptance.
Review #2
- Please describe the contribution of the paper
The article addresses the problem of estimating 3D cell volumes from multi-focus microscopy images. Building the established Deep Image Prior (DIP) method, they propose a deep learning method based on cross-modality distillation to extract 3D cell volumes with significantly increased efficiency. The authors show high quantitative results, while requiring orders of magnitude smaller computation time.
- 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 work addresses a relevant problem. The idea to approximate the expensive reconstruction process and replacing it by an efficient deep learning based approach is sound, and the achieved runtime reduction is significant. The presentation is largely clear, with some minor issues (see below). However, I have several concerns regarding the used methodology and the evaluation, described below.
- 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.
- My greatest concern is regarding the used methodology. While the use of a student teacher model generally appears feasible, it is entirely unclear to me why this relatively complex system of interacting models is necessary and superior over, for example, a simple U-Net model that predicts 3D volumes from multi-focus images end-to-end. I would expect to see the obtained results compared to such simpler models, which could then (potentially) motivate the use of the proposed pipeline. As currently presented, the complex pipeline is not justified.
- Along those lines: The authors recognize the lack of sharpness in the generated results, and attribute it to the probabilistic nature of the VAE (which I would agree on). However, as mentioned above, it is not clear to me why a probabilistic model is used in the first place, if the downstream application (i.e., reconstruction) is fully deterministic and does not require sampling.
- The differences in quality between the generated and DIP results are mentioned (e.g., intensity variations, blurieness), but the effect of these results are not appropriately discussed. Are these deviations from the groundtruth relevant in practice? Are the results usable like this?
- I would also like to raise concerns regarding the use of skip connections in the presented CellVAE architecture. As I understand, the purpose of the VAE is to capture the latent properties of the input. However, the skip connections effectively sabotage this objective, as information can freely flow more or less directly from input to output. Optimization wise, I would deem it counter productive to use skip connections in the VAE.
- I could not find a definition of a training and test set, and it was not specified on which data the metrics were computed. This should be clarified. Metrics computed on the training have little meaning to assess overall performance.
- I would expect more information on the used dataset. Is it an in-house dataset, or was a public dataset used? It would also benefit the paper to briefly explain what NILM, LSIL, HSIL, and SCC are to better understand the difference.
- Fig. 5 and 6 use different color maps. What is the reason for this? Also, difference images in Fig. 4 would make it easier to recognize where the proposed model deviates from the target.
- In 3.1, the authors state: “This means that the two metrics represent the superiority of DeepCVR to the DIP-based method”. This statement is wrong. As the DIP based method is the target, the metric can only measure the alignment between prediction and target, but not superiority.
- The phrasing of the abstract is somewhat clunky and not easy to understand, e.g., “Therefore, the use of the 3D volumes enables to achieve”, or “Here, there is a Deep Image Prior Method”. This should be revisited to state the problem statement and approach more clearly.
- 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.
- 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.
(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 paper addresses a relevant problem and the underlying idea of speeding up the reconstruction is sound. However, due to my concerns regarding the methodology, the lack of simple baselines (e.g., end-to-end U-Net), as well as not enough focus on the quality of the reconstruction (rather than speed), I cannot recommend acceptance.
- 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 thank the authors for their response. However, I am still not convinced that the employed rather complex pipeline is sufficiently justified. While I see the provided explanation of a decomposition into simpler tasks as reasonable, I think the paper requires a quantitative evaluation of appropriate baseline methods to justify this. At the very least, a U-Net baseline would have been necessary. The lack of a proper evaluation of the method components limits it’s transferability to other cases. I thus keep my initial suggestion and recommend reject.
Review #3
- Please describe the contribution of the paper
This paper proposes an algorithm for reconstructing 3D cell volumes from multi-focus image sequences. The proposed DeepCVR utilizes a VAE to learn latent space representations from 3D cell volumes and employs cross-modality knowledge distillation to achieve the mapping from multi-focus images to their corresponding 3D volumes. The core contribution of this work lies in training a student encoder to process cross-modal data, thereby unifying the latent space representations. Furthermore, compared to the DIP-based baseline, the proposed approach significantly accelerates the inference speed, demonstrating practical value for real-world application scenarios.
- 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 methodological details and network architectures are comprehensively and clearly described throughout the paper, which facilitates the reproducibility of the proposed approach. 2.The authors explicitly discuss the computational efficiency and time-cost advantages of the algorithm. This significant reduction in inference time highlights the method’s potential for real-world deployment and clinical translation.
- 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 manuscript requires proofreading to improve its academic tone and formatting. There are instances of colloquial phrasing and obvious grammatical errors, such as the sentence: “Also, the multi-focus image sequence be obtained using the microscope without additional hardware modifications.” These issues should be carefully corrected. 2.Figure 2 lacks an explicit representation of the latent variable s_{3D}, which is discussed in the main text. The figure should be updated to include this annotation for better alignment with the text. 3.The objectivity of the experimental setup is flawed due to a circular dependency in data generation. According to the manuscript, the 3D cell volumes (ground truth) used for training and validating DeepCVR are generated by the DIP-based method. However, the DIP-based method simultaneously serves as the primary baseline for evaluation. This compromises the fairness of the experiments. Ideally, the authors should utilize open-source datasets with established gold standards. In the absence of available objective datasets, ground truths should be simulated using physical models or independently validated third-party methods rather than the baseline method itself. 4.In the experimental section, critical evaluation metrics (PSNR and SSIM) are “calculated by comparing the reconstructed volumes by DeepCVR and the DIP-based method.” Consequently, achieving high metric scores merely demonstrates that DeepCVR’s outputs closely approximate those of the DIP-based method. This mathematical proximity cannot substantiate claims regarding “the superiority of DeepCVR to the DIP-based method,” which further highlights the aforementioned flaw in the dataset construction. 5.Based on the current results, it is evident that DeepCVR achieves higher inference efficiency compared to the DIP-based method. However, there is a lack of valid numerical evidence to demonstrate a distinct advantage in reconstruction performance. If the primary objective of this paper is solely to demonstrate that DeepCVR accelerates inference while strictly maintaining the performance of the DIP-based method, the manuscript’s claims must be revised to accurately reflect and emphasize this specific contribution. Conversely, if the authors intend to claim superior reconstruction performance, they must include additional baseline comparisons (beyond the DIP-based method) or revise the dataset and metric calculation methods to genuinely highlight DeepCVR’s performance improvements.
- 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.
- 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
It is highly recommended that the authors focus their revisions primarily on the dataset construction process and the experimental design. Supplementing and modifying these aspects are crucial to ensure the fairness of the evaluation data and the overall validity of the experimental results.
- 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 proposed method effectively addresses the computational cost of inference and emphasizes the practical usability of the algorithm, demonstrating distinct value for real-world applications. However, these merits are currently offset by shortcomings in the dataset construction and experimental design, rendering the existing results insufficient to fully substantiate the paper’s claims. The manuscript can be considered for acceptance provided that the authors adequately address these critical issues through further discussion during the rebuttal phase or via substantial revisions to the paper.
- Reviewer confidence
Somewhat confident (2)
- [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.
Through the authors’ rebuttal, the motivation and methodology of the proposed work are commendable, and my initial concerns have been addressed. Provided the authors explicitly clarify in the manuscript that the core value of the method lies in the acceleration of Deep Image Prior (DIP), the current work is suitable for publication. Accordingly, I recommend a decision of ‘Accept (Weak)’. However, due to a lack of extensive experience in this specific research direction, I maintain potential concerns regarding two aspects: first, whether the teacher-student architecture remains at the forefront of the field, and second, whether other end-to-end methods have already achieved high-performance and high-efficiency reconstruction, which would imply that DIP, as the primary baseline, may no longer represent the absolute state-of-the-art. Consequently, I request that the meta-reviewer focus on the perspectives of other reviewers regarding these issues to facilitate a more rigorous final judgement.
Author Feedback
We thank all reviewers for their constructive feedback. We address the major concerns below. (1) Justification for VAE + Teacher-Student over End-to-End U-Net (R2, Meta) Our study faces a large cross-modality domain gap between 2D multi-focus images (input) and 3D volumes (output), with limited training data (see (4)). A direct end-to-end model therefore needs to simultaneously learn both cross-modal mapping and 3D volume reconstruction under these constraints. Our strategy for addressing this challenge is to decompose it into two simpler tasks. First, CellVAE (teacher) learns both the latent representation (z_3D) and intermediate features (s_3D) from 3D volumes alone. Second, MF-CellEncoder (student) aligns these intermediate and latent representations. This decomposition is motivated by cross-modal knowledge distillation in [15]. A direct end-to-end U-Net comparison would be a valuable future extension. (2) Why a Probabilistic VAE for Deterministic Reconstruction? / Skip Connections (R2, Meta) During inference, we use the deterministic mean representation μ_MF directly without sampling. The probabilistic formulation is used only during training to regularize the latent space and encourage a smooth and structured representation, facilitating cross-modal alignment through the distillation loss in Eq. 3.Regarding the skip connections, these pathways pass intermediate features s_3D, explicitly aligned with MF-CellEncoder through the third term of Eq. 3, following intermediate-representation distillation in [15]. Without these skip features, the distillation would rely primarily on z_3D. In this case, DeepCVR would potentially lose spatial information critical for high-fidelity decoding. (3) Circular Dependency: DIP as Both Training Target and Baseline (R4, R2, Meta) We acknowledge this limitation. The primary contribution of DeepCVR is to achieve approximately 1.1x10^4 speedup while maintaining reconstruction quality comparable to DIP. However, there is no publicly available ground-truth 3D dataset for bright-field multi-focus cell reconstruction. Therefore, we use DIP-reconstructed volumes based on physical imaging models in [2,13] as the best available pseudo-ground truth. We will revise all related claims, including the imprecise phrase “superiority of DeepCVR to the DIP-based method” in Section 3.1, to accurately reflect this contribution. (4) Train/Test Split Not Specified (R2) We did not use all 320 samples as training data. All samples were split randomly into 80% training (256 samples), 10% validation (32 samples), and 10% test (32 samples). The metrics in Table 1 were computed exclusively on the test set. We will explicitly state this split in the revision to prevent misinterpretation. (5) Blurriness and Latent Space Effectiveness (R1) The consistently high PSNR (avg. 42.99 dB) and SSIM (avg. 0.9974) on the test set indicate that the distillation successfully transfers both intermediate features and latent representations from CellVAE to MF-CellEncoder. The observed blurring is an inherent characteristic of VAE-based smooth latent regularization, as acknowledged in Section 3.3.Per-dimension KL divergence analysis to further characterize posterior behavior would be a valuable future direction. [Minor Points] (R1, R4) (a) The classification of NILM, LSIL, HSIL, and SCC is based on the Bethesda System. (b) We will annotate s_3D and z_3D in Fig. 2.(c) We will clarify that the padded 12th z-slice is discarded prior to loss computation. (d) Fig. 4 shows the reconstructed 3D volumes in which the color of each voxel is determined by its transmittance value, while Fig. 5 shows the grayscale multi-focus images simulated via [13]. We will clarify this in the revision. (e) We will also correct all grammatical errors noted by R2 and R4.(f) Upon acceptance, we will release our code publicly. As our data consists of internal clinical data, public release is not possible.
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 addresses the problem of 3D volume reconstruction from multi-focus images with the goal of speeding up the reconstruction. The proposed model builds on the Deep Image Prior (DIP), here combined with a teacher-student distillation approach to process cross-modal data. The three reviewers highlight the clarity of the exposed method and the important reduction in computational time.
The main raised drawback is the lack of justification for such a complex pipeline (R2): why teacher-student distillation instead of, for instance, an end-to-end U-Net? (not compared to), why a probabilistic model if reconstruction is deterministic and does not need sampling? How to justify the use of skip connections in a VAE architecture? Another important issue is the potential lack of objectivity in the experimental setup (R2, R4), given that DIP is used for training and is thus considered the target. As a result, the current PSNR and SSIM scores cannot support the claim that DeepCVR performs better than DIP (other than in speed), calling for a revision of the claims.
In addition, R1 and R2 note that the reconstructed volumes are blurry. Also, R2 suggests better describing the dataset and experimental setup (train/test sets), the metrics used, and discussing whether the achieved performance deviations are relevant in practice. Finally, although overall the paper was considered clear and reproducible from the explanations, several clarification remarks were made regarding the effective use of latent variables (R1), the treatment of the padded slices in the downstream task (R1), and a need for proofreading (R2 & R4),
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
This paper remains borderline after the clarifications of rebuttal. While R1 and R4 support acceptance focusing on the obtained speedup. R2 advocates for rejection, given the lack of comparison to a straightforward baseline to justify the complexity of the proposed approach, and the reliance on another algorithm DIP used both as target and a baseline (given there is no ground truth at the moment).
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 reviews remain somewhat mixed. R2 is still not fully satisfied with the revision and maintains a recommendation of reject. However, the majority of the reviewers view the revision positively and consider that the authors have adequately addressed the main concerns. Taking the overall discussion into account, I recommend acceptance.
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.
I recommend acceptance as the authors have addressed the main concerns about the latent space design, the motivation for the architecture, and the interpretation of the evaluation. The rebuttal also clarified that the core contribution is the large speedup over DIP while maintaining comparable reconstruction quality. Although minor methodological questions and inextensive experimental varidations remain, they were sufficiently clarified for a positive final judgment, and the overall post-rebuttal feedback supports acceptance.