List of Papers Browse by Subject Areas Author List
Abstract
HoloPointNet presents a novel deep-learning framework for 3D point cloud holography. Generally, computer generated holography (CGH) methods, typically rely on stacked 2D slices and suffer from inefficiencies. These 2D slices often contain empty regions in natural 3D scenes or are intentionally sparse in applications like holographic optogenetics. This results in excessive memory consumption and increased processing latency. In contrast, HoloPointNet directly processes 3D point cloud data using a concatenation-based feature extractor, followed by hierarchical upsampling and wavefront reconstruction modules, eliminating redundant spatial regions and improving efficiency. This design allows for the direct mapping of point cloud data to phase modulations for spatial light modulators (SLMs). By employing a structured convolutional feature transformation pipeline, HoloPointNet enables hierarchical refinement of spatial embeddings, enhancing feature encoding accuracy. HoloPointNet offers the capability to generate multiplane holograms, effectively addressing the complexities of 3D volumetric data. This capability, combined with fast inference times, enables real-time holography for applications such as optogenetics.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4105_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{AmrAnk_HoloPointNet_MICCAI2025,
author = { Amrutkar, Ankit and Nazlioglu, Ahmet and Kampa, Björn and Schulz, Volkmar and Stegmaier, Johannes and Rothermel, Markus and Merhof, Dorit},
title = { { HoloPointNet: A Deep Learning Framework for Efficient 3D Point Cloud Holography } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15970},
month = {September},
page = {267 -- 277}
}
Reviews
Review #1
- Please describe the contribution of the paper
HoloPointNet a novel holography model that process point cloud information into phase encodings then later multiplane holography representation (3D hologram).
- 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.
-
Model design and ablation study is thorough.
-
The HoloPointNet model is novel because it introduces a deep learning architecture that directly maps 3D point clouds to phase encodings for efficient, real-time multi-plane hologram generation—bypassing traditional iterative optimization methods like GS3D.
-
- 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.
It seems the model is heavily inspired by DeepCGH2.0 [9], taking input point cloud. Without direct comparison is extremely hard to determine your “speed” claims and/or increase in other metrics such as Accuracy or Contrast. While they do not have the same outputs, you have clearly adapted 3DGS to fit your evaluation, why not with DeepCGH and 2.0?
- 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 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.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The model has no direct comparison to state of the art or modern techniques in CGH. Whilst the authors claim that the output representation is different they still compare against 3DGS which also has a different representation.
- 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.
I believe the authors could have strengthened their evaluation by including more direct comparisons or failure cases involving DeepCGH and DeepCGH2.0. However, given the fundamental differences in input representations, forward models, and the unavailability of some codebases, I find their explanation reasonable. The choice to compare primarily against a forward model-matched Gerchberg–Saxton baseline provides a fair and principled evaluation. Therefore, I am satisfied and recommend acceptance.
I believe this work represents a smart architectural design and progress towards practical, unsupervised neural holography that accurately models free-space propagation for multiplane 3D reconstruction.
Review #2
- Please describe the contribution of the paper
This work proposes a deep learning framework that uses point cloud representations instead of traditional voxel representations, reducing redundant data representations in holographic optogenetics and achieving real-time performance in one holographic optogenetics application.
- 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 evaluation results is interesting. Table 1 and Fig.3 clearly demonstrates the performance of proposed method in this paper outperforms its baseline and a state-of-the-art work.
- 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.
Authors should clarify the background, challenges in this research field and their motivations in Introduction.
And the authors only compare with two state-of-the-art methods (Baseline and 3DGS) and don’t cite them clearly in Results, which is inadequate in my opinion.
- 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
-
Authors should explain ‘phase retrieval problem’ explicitly in the Introduction if this is a main problem that has been addressed in this paper. If not, please remove redundant information and retain key information to claim your motivations, like ‘memory consumption’ and ‘inference times’.
-
The proposed deep learning network seems overly simple and using point cloud as representations is not novel.
-
If ‘3D Gerchberg-Saxton’ originates from a scientific paper, please cite it.
-
It would be better to describe your ablation studies (1,2,3,4) in a table.
-
Please add citation and/or name of the Baseline in your figures and tables. Is GS3D you mentioned ‘3D Gerchberg-Saxton’ 3DGS? If so, please keep the name consistent, and cite the original paper.
-
You’ve already shown the qualitative results of Baseline, GS3D, and ablation studies in Fig.2, why not show the results with your full model?
-
Can you add citations of the ‘standardized test dataset’ if possible?
-
What is the unit of Time in your Table 1?
-
- 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?
- Authors only compare with two methods, one is ‘Baseline’, another is ‘3DGS’, but not mention their references.
- The proposed method (deep learning network) seems overly simple and using point cloud as representations is not novel.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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’ rebuttal addresses my concerns and clarifies the novelty of their work to some extent. I believe only minor revisions are needed before the paper is suitable for publication.
Review #3
- Please describe the contribution of the paper
Authors propose a neural network that directly maps point clouds to phase encoding for efficient multiplane holography. Unlike conventional computer-generated holography (CGH) methods, which rely on inefficient 2D slices, their approach encodes only essential target locations, thus reducing memory and latency, which is critical for sparse applications like holographic optogenetics. Their concatenation-based features extractor, hierarchical upsampling, and wavefront reconstruction modules facilitate fast, real-time hologram generation. Results demonstrate improved accuracy and faster inference compared to the adapted GS3D model. The method is computationally efficient, scalable, and has been successfully extended to 2D free-space and Fourier holography, showcasing its versatility.
- 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 presents a novel point cloud-based representation for computer-generated holography (CGH), moving beyond traditional dense 2D slice representations. This formulation is particularly efficient for sparse 3D scenes and applications such as holographic optogenetics. The proposed neural network architecture, featuring a concatenation-based feature extractor, hierarchical upsampling, and a dedicated wavefront reconstruction module, enables direct mapping from sparse inputs to phase-encoded holograms with reduced computational overhead. The model demonstrates real-time inference capability, making it suitable for applications requiring rapid holographic updates, such as closed-loop neural stimulation. Both qualitative and quantitative evaluations show that the method outperforms the adapted GS3D iterative algorithm in terms of phase reconstruction accuracy, contrast, and computational efficiency. The versatility of the model is further demonstrated by its successful adaptation to 2D holography settings, including free-space propagation and Fourier holography. Additionally, by encoding only essential target points, the method significantly reduces memory consumption and latency, offering a scalable and resource-efficient solution for CGH.
- 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.
Although authors demonstrate robustness by evaluating their method, HoloPointNet on simulated CGH datasets and achieving notable metrics, several limitations exist. The qualitative analysis of Ablation study 4 highlights improved identification of high-amplitude regions, but there is no discussion regarding phase information after phase reconstruction. Moreover, while Ablation 4 outperformed others in accuracy, efficiency, and contrast, the speckle contrast was significantly higher, indicating poorer performance, which was not addressed anywhere in the paper. Authors also did not discuss any limitations of their approach. Furthermore, the model was evaluated on only simulated images using simple circular points, which lack the complexity and heterogeneity of real biological structures such as neurons. The hologram reconstructions shown in Fig. 2 do not display much detail, and no phase retrieval information is presented. Evaluating the model on a more complex simulated or biological dataset would strengthen the credibility and help assess the broader clinical applicability of the method. Nevertheless, despite these limitations, the proposed approach introduces a novel framework for encoding point clouds into phase representations for holographic reconstruction. The method demonstrates promising results on synthetic datasets and suggests potential for future real-time holographic applications, especially in optogenetics. Addressing the identified weaknesses, particularly through evaluation on more complex or biological datasets and deeper analysis of phase quality, would further strengthen the impact and clinical relevance of the work.
- 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.
- 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 should revise the in-text citation numbering to ensure that all citations appear in the correct sequential order. In Fig. 1, a clear and descriptive figure caption must be provided to explain the figure contents. The current numbering of weighted equations is confusing; for instance, equation (1) and weight label (2a) are not distinguishable. Authors should revise the notation to avoid ambiguity between equations and weights. In Fig. 2, the comparative studies between the proposed method, the baseline, and the GS3D models are difficult to interpret. Authors should explicitly indicate which column corresponds to their proposed method. Additionally, Table 1 should be properly formatted with a table heading, rather than embedding it within the caption.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper proposes a novel method that significantly reduces computational overhead compared to traditional grid-based CGH methods and shows promising performance on simulated datasets. However, it is evaluated only on simple simulated images that do not fully capture the complexity of biological structures, and there is limited discussion regarding the quality of phase retrieval and speckle noise performance. While these limitations restrict immediate clinical applicability, the conceptual innovation, efficient design, and strong baseline results support a strong score, as the method holds high potential for future impact once extended and validated further.
- Reviewer confidence
Very confident (4)
- [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.
While the authors have acknowledged the limitations and expressed a willingness to address them in future revisions, the current version of the paper leaves several critical concerns unresolved. The method is evaluated solely on simplified simulated CGH data using basic circular structures, which do not reflect the complexity or heterogeneity of realistic biological samples. As a result, the generalizability and applicability of the proposed approach remain uncertain. Furthermore, although Ablation Study 4 is reported to improve accuracy and amplitude recovery, the authors themselves note that it produces significantly higher speckle contrast, an indicator of degraded reconstruction quality. This trade-off is neither discussed nor analyzed, and the phase retrieval aspect of the holographic reconstruction is not addressed at all. The qualitative images presented in Fig. 2 do not provide convincing evidence of improved reconstruction, and the lack of phase analysis undermines the overall contribution of the method. Given that these issues were not meaningfully addressed in the original submission or the rebuttal, and considering the evaluation is limited to a narrow synthetic setup, the paper is not ready in its current form. I encourage the authors to strengthen the work by incorporating more realistic datasets, a deeper analysis of phase information, and a balanced discussion of the method’s limitations.
Author Feedback
We thank all reviewers for their constructive feedback and are happy to clarify some of the raised concerns in this letter.
On performed comparisons (R3 + R4): Our approach uses free-space propagation (FSP) as the forward model, where multiple hyperparameters play a significant role in the neural network performance (SLM size, pixel pitch, wavelength, and propagation distance).Typical models are designed and trained by keeping the forward model hyperparameters (FMH) fixed and require training a new network upon changing the FMH (or adapt the network design) to get better results [1]. In contrast, existing models such as DeepCGH (open-source) and DeepCGH2.0 (not open- source) use Fourier holography as the forward model, where only SLM size is typically fixed. DeepCGH supports multiplane holography with grid-based inputs and DeepCGH2.0 uses point cloud inputs, but does not support multiplane holography. Neither of these models were designed for FSP with multiple fixed hyperparameters. As both the input representations and forward model assumptions differ, adapting DeepCGH variants to a FSP setting would require substantial changes and may not yield meaningful comparisons. To the best of our knowledge, there is no “unsupervised” CGH model using a FSP forward model with “point cloud” inputs for “multiplane” reconstruction available for direct comparison. Since our method, like many CGH approaches, can be interpreted as an unrolled version of the Gerchberg–Saxton (GS) algorithm [1][2], we chose to compare against a GS algorithm adapted for FSP with matching forward model hyperparameters. This provides a consistent and principled basis for evaluation. We will add additional explanations for the baseline selection and limitations of existing approaches.
On the model (R4 + R5): Baseline and Ablation 4, both are variants of the proposed HoloPointNet. The difference lies in the respective Wavefront Reconstruction modules. As depicted in Fig. 1, the Wavefront Reconstruction module (light red) has a hierarchical channel reduction for the Baseline model and abrupt channel reduction for the Ablation 4 model (Sec. 2.3 and Sec. 4). We will extend the figure caption and explanations to clarify this difference.
On novelty (R4): To the best of our knowledge the presented work is the first deep learning-based model for multiplane holography that combines free-space propagation with memory-efficient point cloud input data.
On textual improvements (R4 + R5): We will revisit the explanations of the phase retrieval problem in the introduction and will formulate it clearer. In particular, we have adapted the GS algorithm by changing its forward model from Fourier holography to free space propagation. The “standardized test data” meant “fixed test data” for all the compared models and we will adjust the text accordingly. In Tab. 1, time is in seconds. We use violin plots so that we can see the spread of the distribution of metrics properly (Sec. 3.) Thanks for spotting some inconsistent specifiers, which will be corrected.
On data (R5): Since our main application is optogenetics, circular blobs mimicking cell bodies are sufficient. We agree testing on more complex data is interesting and leave it for future work.
On phase retrieval information (R5): The quality of the output images and the quantitative metrics reflect the quality of the phase mask that the network approximates at the SLM plane (Eq. 2b).
On Ablation 4 (R5): The poor performance of Ablation 4 in terms of speckle contrast is mentioned in the caption of Fig 3 and we will add a brief discussion of this effect.
[1] A. Amrutkar, et al. “Towards Robust and Generalizable Gerchberg Saxton based Physics Inspired Neural Networks for Computer Generated Holography: A Sensitivity Analysis Framework.” arXiv preprint arXiv:2505.00220 (2025).
[2] R. W. Gerchberg, “A practical algorithm for the determination of phase from image and diffraction plane pictures.” Optik, 35, 237-246., 1972
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”.
N/A
- 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’
There remains a divergence of opinions among the reviewers after the rebuttal phase. However, the supportive assessments from R3 and R4 appear well-founded, especially after the authors provided a convincing and thorough rebuttal. While R5 initially gave a positive evaluation, their post-rebuttal comments introduced new concerns that were not raised in the first-round review, leaving the authors no opportunity to address them appropriately. As these additional critiques would require substantial further analysis and likely another review cycle, I consider it fair to set them aside for this decision. Based on the overall evaluation and the strength of the rebuttal, 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A