List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Jiaqi Zhang, Yan Hu, Xiaojuan Qi, Ting Meng, Lihui Wang, Huazhu Fu, Mingming Yang, Jiang Liu
Abstract
The shape of the posterior eyeball is a crucial factor in many
clinical applications, such as myopia prevention, surgical planning, and
disease screening. However, current shape representations are limited by
their low resolution or small field of view, providing insufficient infor-
mation for surgeons to make accurate decisions. This paper proposes
a novel task of reconstructing complete 3D posterior shapes based on
small-FOV OCT images and introduces a novel Posterior Eyeball Shape
Network (PESNet) to accomplish this task. The proposed PESNet is de-
signed with dual branches that incorporate anatomical information of
the eyeball as guidance. To capture more detailed information, we intro-
duce a Polar Voxelization Block (PVB) that transfers sparse input point
clouds to a dense representation. Furthermore, we propose a Radius-wise
Fusion Block (RFB) that fuses correlative hierarchical features from the
two branches. Our qualitative results indicate that PESNet provides a
well-represented complete posterior eyeball shape with a chamfer dis-
tance of 9.52, SSIM of 0.78, and Density of 0.013 on the self-made pos-
terior ocular shape dataset. We also demonstrate the effectiveness of our
model by testing it on patients’ data. Overall, our proposed PESNet
offers a significant improvement over existing methods in accurately re-
constructing the complete 3D posterior eyeball shape. This achievement
has important implications for clinical applications.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43987-2_18
SharedIt: https://rdcu.be/dnwJA
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #3
- Please describe the contribution of the paper
The paper describes an approach to create a voxelized representation of the eyeball shape in polar coordinates based on a number of limited field-of-view OCT scans, using a dedicated trained network.
- Please list the main strengths of the paper; you should write about 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 addresses an existing clinical need in an original manner. The method is evaluated on 55 scanned eyes, that are split into training, validation and test data in the ratios 70%, 20%, 10%. The method is favorably compared to a reimplementation of a state-of-the-art method (PointNet).
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
The data set (used for training and evaluation) consists of healthy volunteers. Thus it remains difficult to assess how well the method serves the clinical needs. Also, it is not clear which performance in clinical practise is needed, and whether this is met or not.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
While the authors explain the overall system, the steps taken and most important aspects, the reproducibility of their work can be called ‘fair’ at best. Without making the data set and the code public, a lot of implementation details have to be reinvented by anyone who wants to reproduce the results.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
- Please briefly explain what the advantage of being able to use several limited FoV OCT over large FoV OCT is
- Methods: please explain (B,C,R,U,V) on their first usage.
- Methods: it is unclear whether the reduction of R to 1 is also performed on the ground truth data (i.e. single surface). If yes, then this may bias your evaluation.
- Experiments: You may consider to move “Implementation Details” to the “Methods” section
- Experiments: 2) L_perc can achieve sub-optimal results -> I suspect that you mean to say that it can achieve very granular results?
Minor details:
- Introduciton: the resolution of MRI scans of the eye will vary between hospitals and imaging vendors. The statement in the introduction is too absolute, and probably refers to the practice in the authors’ institution
- Voxels are volumetric elements, and therefore their unit would be mm^3, not mm^2
- Please be consise in the symbol used for multiplication. Now the first time you use the cross symbol, and later the asterix
- Table 2: please use the same L font for L_ssim
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Strongest driver in the recommendation was the inability to assess to which degree the method described in the paper has fulfilled the clinical needs.
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #2
- Please describe the contribution of the paper
• The authors proposed a new task to reconstruct the complete posterior eyeball shape within large FOV only based on local OCT images. • The authors developed a novel Posterior Eyeball Shape Network (PESNet) leveraging the anatomical prior of the eyeball structure to accomplish the above task. • The authors tested the PESNet qualitatively and quantitatively on self-made posterior ocular shape dataset as well as on patients’ data.
- Please list the main strengths of the paper; you should write about 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 studied problem is new and of high significance. • The proposed method is innovative and technically sound. • The empirical study on self-generated data and real-world patient data yielded promising results. • The manuscript is well-structured.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
• The neural network structure is quite complicated. It would be helpful if the rationale behind the design could be better explained. • The author stated, “Given the center coordinate c = (cx, cy, cz) of the top en-face slice of OCT images as polar origin, the transformation between the cartesian coordinate (x, y, z) and the polar coordinates (r, u, v) can be expressed as Equation 2.” => However, it is unclear how to make sure the origin of the subarea point cloud aligns with that of the template point cloud. It seems there is no sufficient information on the subarea point cloud to locate the correct global origin. • The author stated, “The inputs are transformed to the polar grid with size (B,C,R,U,V) by” => It is unclear where to find the definition / explanation of the valuable B. • The empirical comparison was primarily an ablation study. The only other method tested is PointNet. It would be helpful to include additional competing methods if there exists anymore. • The tested data sets are small.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
• While there is a lack of some technical details, the supplementary materials provide additional information to help improve the reproducibility.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
• The neural network structure is quite complicated. It would be helpful if the rationale behind the design could be better explained. • The author stated, “Given the center coordinate c = (cx, cy, cz) of the top en-face slice of OCT images as polar origin, the transformation between the cartesian coordinate (x, y, z) and the polar coordinates (r, u, v) can be expressed as Equation 2.” => However, it is unclear how to make sure the origin of the subarea point cloud aligns with that of the template point cloud. It seems there is no sufficient information on the subarea point cloud to locate the correct global origin. • The author stated, “The inputs are transformed to the polar grid with size (B,C,R,U,V) by” => It is unclear where to find the definition / explanation of the valuable B. • Too many acronyms were used in the manuscript. It would be helpful to provide an acronym table so that the readers could easily find their full descriptions as needed. • The empirical comparison was primarily an ablation study. The only other method tested is PointNet. It would be helpful to include additional competing methods if there exists anymore. • It would be more convincing if more data sets were tested in the empirical study.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
• The problem is new and of high importance. • The method is new and technically sound. • The empirical study can be improved.
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #1
- Please describe the contribution of the paper
This paper proposes the task of reconstructing complete 3D posterior shapes based on small-FOV OCT image, and introduces a novel network for this purpose, Eyeball Shape Network (PESNet). The result is a 2D ocular surface map (OSM). The network includes various branches including shape regression, anatomical prior, fusion block, a polar grid arrangement is used due to the approximately spherical shape. A ground truth dataset of 55 eyes is generated. Results provide chamfer distance (CD) and Structural Similarity Index Measure (SSIM) as 2D metrics to evaluate the OSM regression performance, and a point density metric evaluates the number ratio of total points between predicted and GT point clouds.
- Please list the main strengths of the paper; you should write about 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 task of reconstructing complete 3D posterior shapes based on small-FOV OCT images is relatively novel (however there seens to be missing related work). The formulation based on polar coordinates is well-suited to the spherical eyeball shape.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
The work addresses a niche domain, there appears to be some missing related literature. The difficult of the task is unclear, it seems there is strong prior regarding the eyeball shape.
- Please rate the clarity and organization of this paper
Very Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
Reproducible.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
A baseline comparison is provided against PointNet[20], a generic method not optimized for spherical 3D shapes, based on the authors own implementation, results are very poor (Table 1). This appears to be a bit of a weak comparison.
The authors claim they are the first to attempt the task of 3D ocular shape reconstruction from OCT, whereas a variety of closely related work has looked at 3D correction of OCT images and should be cited. Specifically reconstruction of the retinal surface and retinal pigment epithelial (RPE) layers as in the current work:
Kuo, Anthony N., et al. “Correction of ocular shape in retinal optical coherence tomography and effect on current clinical measures.” American journal of ophthalmology 156.2 (2013): 304-311.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The task seems novel.
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Primary Meta-Review
- Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. 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.
This paper proposed the task of reconstructing complete 3D posterior shapes based on small-FOV OCT image, and introduces a novel network for this purpose, Eyeball Shape Network (PESNet). Experiments were performed 55 eye images with groundtruth data. Although the reviewers identified some weakness, such as missing literature and limited dataset size, they all appreciated the novelty of the work and showed strong enthusiasm. The authors may incorporate the reviewer comments in their camera ready version.
Author Feedback
We thank the reviewers for high-quality comments. Below we provide point-to-point responses to the comments, which will be integrated in the final version. [Q] The baseline comparison with PointNet. (R1) [A] To make PointNet suitable for our task, we modified it based on its original segmentation network to take local point cloud as input and output point coordinates of global point cloud, and adjusted the size of MLP layers to maintain acceptable model size. Besides, we perform special preprocessing for input and GT data including down-sampling 4096 points at regular intervals, to preserve as much the shape information of RPE layers as possible. More details are listed in the supplementary material. Its poor results may be mainly due to the limited number of input points (4096), losing most of structural information and surface details. Compared to the original GT point cloud containing 76.8k points, 4096 points are too sparse to provide enough structural information. [Q] Additional competing methods. (R2) [A] To the best of our knowledge, there are currently no other methods available for direct comparison tailored to our novel task. For advanced 3D reconstruction method in natural scene, NeRF- and MVS-based methods require multi-view images and camera ex- and intrinsic parameters as input which can’t be obtained by ophthalmic imaging device at all. Different from point cloud completion task that given global structure to fill local details, our task is to reconstruct global shape based on local information. For common CNN-based methods, it’s hard to trade-off between memory overheads and volume size, especially for sparse data. And its reconstructed results need complex and non-differentiable post processing for both voxel- and grid-based data format. [Q] Add an extra literature about 3D correction of OCT images. (R1) [A] Okay, we will add this literature in our reference. This work is to correct the curvature of OCT volume in the scan FOV, to recover the flatten distortion caused by A-scans being plotted in a parallel format in some OCT devices. In contrast, the ultra-widefield OCT device we used can obtain the scan volume with accurate curvature, so there is no need to perform extra correlation. Besides, our work is to reconstruct the global eyeball shape based on local OCT, whose FOV is beyond that of local OCT scan. So, there are differences between the two tasks. [Q] Add the definition of the valuable B of (B, C, R, U, V). (R2/R3) [A] We will add this definition in full version paper. The valuable “B” in (B, C, R, U, V) denotes the batch size, the other valuables have been introduced in paper. [Q] Did the reduction of R to 1 also performed on the ground truth data? (R3) [A] Yes, we performed the reduction of R on the ground-truth grid to make up 2D GT-OSM. We also evaluated the results from both 2D and 3D perspectives using both Chamfer Distance, SSIM and Density, to minimize any potential biases. [Q] May consider moving “Implementation Details” to the “Method” section. (R3) [A] Thank you for your suggestion. We will consider revising it. [Q] “sub-optimal” meaning. (R3) [A] The term ‘suboptimal’ describes that the L_perc performs below our combined loss function in Table 2. The performance of using the L_perc alone still obviously inferior to our combined Loss. [Q] Add more data. (Meta/R2) [A] Yes, we are collecting more data including healthy and diseased cases to enlarge our datasets. [Q] Add an acronym table and rationale behind the design of network structure.(R2) [A] We will add acronym into corresponding figures and rationale into paper if there is space left. Our design uses polar transformation to tackle RPE label sparsity and meanwhile maintain acceptable size of CNN model. PVB first covert cartesian point clouds to polar grids, densifying sparse data. Then, anisotropic convolution condenses C and R dimensions, converting the complex 3D reconstruction into a simpler 2D regression task, downsizing the model.