List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Marcel Beetz, Abhirup Banerjee, Vicente Grau
Abstract
Myocardial infarction (MI) is one of the most common causes of death in the world. Image-based biomarkers commonly used in the clinic, such as ejection fraction, fail to capture more complex patterns in the heart’s 3D anatomy and thus limit diagnostic accuracy. In this work, we present the multi-objective point cloud autoencoder as a novel geometric deep learning approach for explainable infarction prediction, based on multi-class 3D point cloud representations of cardiac anatomy and function. Its architecture consists of multiple task-specific branches connected by a low-dimensional latent space to allow for effective multi-objective learning of both reconstruction and MI prediction, while capturing pathology-specific 3D shape information in an interpretable latent space. Furthermore, its hierarchical branch design with point cloud-based deep learning operations enables efficient multi-scale feature learning directly on high-resolution anatomy point clouds. In our experiments on a large UK Biobank dataset, the multi-objective point cloud autoencoder is able to accurately reconstruct multi-temporal 3D shapes with Chamfer distances between predicted and input anatomies below the underlying images’ pixel resolution. Our method outperforms multiple machine learning and deep learning benchmarks for the task of incident MI prediction by 19% in terms of Area Under the Receiver Operating Characteristic curve. In addition, its task-specific compact latent space exhibits easily separable control and MI clusters with clinically plausible associations between subject encodings and corresponding 3D shapes, thus demonstrating the explainability of the prediction.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43895-0_50
SharedIt: https://rdcu.be/dnwy3
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #2
- Please describe the contribution of the paper
The paper proposes a new deep-learning architecture to predict the myocardial infarction based on point cloud autoencoders. The strategy sounds interesting by outperforming other simple artificial intelligence models
- 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 point cloud autoencoders provides 3D anatomical shape information for a more precise characterization of the LV.
- 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.
A missing comparison with other state-of-the-art methods. The paper shows only results for machine learning model and deep learning models
- 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
The paper can be reproduced
- 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 authors should include more state-of-the-art methods for myocardial prediction to validate the model performance. The method can detect little patterns of myocardial infarction, this is useful for a more precise detection.
- 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 method sounds interesting, but the performance is not good validated. There a lot of methods that are focused on myocardial prediction. Such that this work should include a good literature review and comparison
- Reviewer confidence
Very confident
- [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 #4
- Please describe the contribution of the paper
The authors proposed in this work a multi-objective autoencoders for myocardial infarction prediction. The model is adapted for point cloud data and able to handle full 3D cardiac anatomy. The authors shows that the network is able to reconstruct accurately the 3D cardiac anatomy and improve the classification performance compare to SOTA methods. An analysis of the latent space using Laplacian eigenmaps is conducted.
- 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 is well written
- Use of a public dataset
- Combine as input for each case at end diastole and end systole and the three anatomical structures in full pipeline
- Improve classification performance compared to SOTA methods
- 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.
Major:
An explainable myocardial prediction furnishes insight of the decision-making process of the prediction. The Fig. 4 shows that the method is able to complex handle variation of cardiac structure, but do not provide explainability. A extensive study of the latent space (variation modes, study of the boundaries in the latent space) or using SHAP values [1] can provide true explainability of the prediction. Also, in a classification problem, it can be expected than the first two dimensions are able to separate the two classes. In an unsupervised setting, this analysis can be meaningful because it reveals insight of the distribution of the data.
[1] https://shap.readthedocs.io/en/latest/index.html
Minor:
-
In this introduction, it is stated “while simultaneously predict future cardiac event” as an objective. The experiment conducted is about predicting previous cardiac events.
-
The quality of the reconstruction is not the main objective, but there is no comparison with other methods, e.g. with or w/o cloud dense branch, and only chamfer distance. Hausdorff distance can also work here
-
Why not compare with the same prediction model (based on MLP) in the w/o multi-objective experiment and use a regression instead.
-
- 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
Good reproducibility of the paper
- 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
Text on supplementary materials not allowed
- 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?
This multi-objective proposed method is able to handle the full 3D cardiac anatomy at two cardiac, perform accurate reconstruction and improve the classification performance. Despite some concern of the explainability of the prediction, the analysis of the latent step proposed is an interessant first step to present to the community.
- 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
The authors propose a multi-objective DNN that reconstructs 3D LV/RV point clouds, and predicts myocardial infarction (MI). Input to the encoder is a vector containing ED/ES point clouds with cardiac structure labels. The encoder maps to a latent space of mean and std (similar to VAE). The encoder is followed by a reconstruction branch and an MI prediction branch. The DNN is trained with a loss that consists of 3 terms : Course and dense PC reconstruction loss based on Chamfer distance, KL divergence between latent space and normal distribution, binary CE loss for the MI prediction. Cine MRI images from the UK Biobank dataset are used for training and validating the method (470 subjects - 50% MI, 50% normal). The shape reconstruction error is shown to vary from 1.26-1.71mm, which are less than the pixel resolution. The MI classification metrics for the proposed method are shown to outperform the author proposed baselines.
- 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 is well written and easy to understand. It addresses a clinically important problem of myocardial infraction prediction. The authors validate their method on a publicly available dataset.
- 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.
-There is very less discussion on prior art for MI prediction using 2d images and 3d pointclouds (especially on prior network architectures) -Point cloud reconstructions are smoothed in the prediction (fig 2) and not always accurate when observed carefully. -Is “p” equal to “n” in the network architecture? If not, is there any reason for this choice? -Table 2 does not have any baseline methods from research literature. How does the proposed method compare against [11, 16]? What about 2D image based methods? [2, 13, 15, 19, 21, 28, 29, 33]
- 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
The authors provided details about the data and cases selected from the UK Biobank. They also provide hyperparameters of the DNN used during training.
- 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 proposed method is an extension to a simple 3D pointcloud VAE with an additional branch from the latent space for MI classification. The validation of the method could have been better with better baselines. There is some merit to using a monotonic annealing schedule during training for the weighting parameter alpha. Although Fig 3 shows some separation between the two classes, the boundary is definitely not distinctive enough (Accuracy of 0.694 and F1-score of 0.695). Hence, the author’s claim of “…offering a high degree of pathology-specific separability …” is inflated and needs to be toned down.
- 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?
This is a well-written paper where the validation of the method could be improved by comparing against other methods reported in literature for MI prediction. A stronger validation would be needed for deployment of such a system in clinical settings.
- Reviewer confidence
Very confident
- [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.
The authors propose a multi-objective DNN that reconstructs 3D LV/RV point clouds, and predicts myocardial infarction (MI). Input to the encoder is a vector containing ED/ES point clouds with cardiac structure labels. The encoder maps to a latent space of mean and std (similar to VAE). The model is adapted for point cloud data and able to handle full 3D cardiac anatomy. The authors shows that the network is able to reconstruct accurately the 3D cardiac anatomy and improve the classification performance compare to SOTA methods
Strengths of the paper:
- The point cloud autoencoders provides 3D anatomical shape information for a more precise characterization of the LV.
- The paper is well written and easy to understand.
- It addresses a clinically important problem of myocardial infraction prediction.
- Combine as input for each case at end diastole and end systole and the three anatomical structures in full pipeline
- The authors validate their method on a publicly available dataset.
- Improve classification performance compared to SOTA methods
Weaknesses of the paper:
- The paper could benefit form a comparison with other state-of-the-art methods (not machine learning model or deep learning models)
- The literature review on prior art for MI prediction using 2d images and 3d pointclouds can be extended.
- It would be nice to add a discussion on the smothering of the point cloud reconstructions and it’s effect on the accuracy of the model. smoothed in the prediction (fig 2) and not always accurate when observed carefully.
- It would be interesting to add some explainability strategy of the prediction.
- Although the quality of the reconstruction is not the main objective, it would be nice to add a comparison with other methods
Recommendation: The authors proposed a novel method to predicts myocardial infarction. The paper would be a good contribution to the MICCAI community but it could be improved by adding some of the reviewers suggestions.
Author Feedback
N/A