List of Papers Browse by Subject Areas Author List
Abstract
We present a novel method for quantifying the microscopic structure of brain tissue.
It is based on the automated recognition of interpretable features obtained by analyzing the shapes of cells. This contrasts with prevailing methods of brain anatomical analysis in two ways.
First, contemporary methods use gray-scale values derived from smoothed version of the anatomical images, which dissipated valuable information from the texture of the images.
Second, contemporary analysis uses the output of black-box Convolutional Neural Networks, while our system makes decisions based on interpretable features obtained by analyzing the shapes of individual cells.
An important benefit of this open-box approach is that the anatomist can understand and correct the decisions made by the computer.
Our proposed system can accurately localize and identify existing brain structures.
This can be used to align and coregistar brains and will facilitate connectomic studies for reverse engineering of brain circuitry.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/2587_paper.pdf
SharedIt Link: pending
SpringerLink (DOI): pending
Supplementary Material: https://papers.miccai.org/miccai-2024/supp/2587_supp.pdf
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Qia_Towards_MICCAI2024,
author = { Qian, Kui and Qiao, Litao and Friedman, Beth and O’Donnell, Edward and Kleinfeld, David and Freund, Yoav},
title = { { Towards Explainable Automated Neuroanatomy } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15003},
month = {October},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents a framework for brain region segmentation from microscopic images using cell shape features at the region level. Currently deep learning approaches already dominates the microscopic image analysis field, while this work develops a classical method with explainable features. The features used in this study are learned unsupervisely plus hand-crafted. The results showed that this work outperforms the CNN approach and the classifier performance is highly explainable. The major contribution of the paper is the extraction and usage of explainable cell shape features which gains comparable results with deep learning approaches.
- 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.
This works uses a conventional classification pipeline, which is explainable and robust against varying staining and image appearances. The demonstration of the probability maps in the result section is also a bonus point.
- 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.
Some of the results seems confusing. Fig.4(b) shows that for most regions this method outperforms CNN (with only two exceptions, 6N and SNR), but the main text says the average AUC of this method is lower than AUC. This is quite conflicting. The method is not compared with the SOTA deep learning frameworks, e.g. the transformer based which might be better than CNN.
- 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.
- Do you have any additional comments regarding the paper’s reproducibility?
N/A
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
The explainable nature of this method is an advantage, but whether this method outperforms the SOTA method in terms of segmentation accuracy is unclear. I would suggest the authors to compare their approach with SOTA method, especially in terms of explainability and robustness, which are the strengths of this method.
- 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
Weak Accept — could be accepted, dependent on rebuttal (4)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The explainable nature of this method is a contribution of this work.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
This work presents a novel method for quantifying the microscopic structure of brain tissue. The proposed system can accurately localize and identify existing brain structures, with explainability for anatomist.
- 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 system utilizes interpretable cell shape features for structure detection, providing a clear and understandable approach.
- The method’s simplicity is a strength, making it accessible and efficient.
- 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.
- It is unclear whether each cell image contains only one cell, which may impact the method’s efficacy.
- The feature crafting process could benefit from an ablation study to understand its contribution to the method’s performance.
- The definition of a region’s boundary and its impact on different region sizes are unclear.
- It is unclear how the classification score reflects the quantity ratio between different structures within a region.
- With only three brain images available, cross-validation is crucial to minimize the impact of a single testing set split.
- A comparison of the proposed method’s computational overhead and running time with CNN would be beneficial.
- 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.
- Do you have any additional comments regarding the paper’s reproducibility?
N/A
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
- Breaking up Section 2 into more paragraphs would improve readability and ease of understanding. Especially paragraph 1 and 2 in Section 2.
- It took time to understand that each feature has a CDF curve discretized into 99-dimensional vectors; additional explanation would be helpful.
- Adding a grid and increasing the offset from 0.4 in Figure 4b would enhance clarity.
- The meaning of “Z” in Figure 5 should be explained in the caption to ensure reader understanding.
- 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
Weak Accept — could be accepted, dependent on rebuttal (4)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The work present an explainable, light-weighted yet effective method to recognize the brain structure, which is pretty novel and in the interests of MICCAI community.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 #3
- Please describe the contribution of the paper
The paper proposes interpretable shape representation of brain cells. Further, the paper proposes a method for structure detection, which can be robust against noise in extracting individual shapes. Given the selected features, the authors adopt XGBoost for classification, where the results can be interpretable. The results suggest the proposed methods yield good accuracy.
- 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 presented, including the methods, the experiments and the results. It is easy to follow the paper.
- The proposed methods are novel. The features selected can be analyzed in an interpretable way.
- The work are motivated by important clinical problems. The design of the methods is closely related to the domain expertise, making the results useful and practical.
- 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.
- It is not clear that Fig 2 chose 1st and 4th eigenvectors to visualize the features.
- In the introduction section page 2, it is not clear what are compared in the statement “more accurately represent the diversity of cell shapes”. Similarly, it is not clear what are compared in the last sentence of the introduction: “making it more resilient to”
- In section 2 last paragraph 1st sentence, it is not clear how the authors “retrieve features for all cells”. Is it based on Euclidean distance of feature vectors?
- Please rate the clarity and organization of this paper
Excellent
- 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.
- Do you have any additional comments regarding the paper’s reproducibility?
NA
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
- As compared to affine transformation of cells, Procrustes alignment can also yield shape regardless of sizes of objects.
- In analyzing shape features, proper distance metric can improve the performance of classifiers that are used in the paper, see the paper “Skeletons, object shape, statistics” for more details.
- 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
Accept — should be accepted, independent of rebuttal (5)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper shows innovative shape representation of brain cells and of brain structures. The methods and the outcome are of clinical use. In general, the paper clearly presented the methods and discussed the results.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
Author Feedback
We appreciate the reviewers’ insightful suggestions.
Reviewer 1.
- It is not clear that Fig 2 chose 1st and 4th eigenvectors … .
We chose to use the 1st and 4th for display purposes. This pair provides a better geometrical interpretation than the 1st and 2nd.
- In the introduction section page 2, [clarify] “more accurately represent the diversity of cell shapes”.
This statement is now corrected to read “… more accurately represent the diversity of cell shapes than prior manual-designed features proposed in reference [9].”
- Similarly, [clarify] “making it more resilient to”.
The statement is corrected to read “… making it resilient to variations in staining methodology and imaging modality.”
- In section 2 last paragraph 1st sentence, [clarify] “retrieve features for all cells”.
This statement makes a qualitative reference to the cell feature vector for each cell. We corrected the text to read “To compute the region feature for a region containing a group of cells, we first query our database to retrieve ‘cell feature’ vectors for all cells within the region, followed by generating cumulative CDFs for these features.”
- Is it based on Euclidean distance of feature vectors?
No. We do not need to query based on distance because all cell feature vectors are stored in the database.
Reviewer 3.
- … Fig.4(b) shows that for most regions this method outperforms CNN …, but the main text says the average AUC of this method is lower than AUC… .
We apologize for any confusion. While our method generalizes for structure detection by fluorescent microscopy, the CNN does not (Fig. 5). However, the CNN performs slightly better in detecting structures on material prepared with the same stain as the training set; note higher CNN (orange) levels vs our method (blue) in Figure 4b. We now state, “The average ROC AUC score for our method is 0.89, which is only slightly lower than the CNN’s average of 0.92 yet still represents a robust performance.”
- The method is not compared with the SOTA deep learning frameworks, … .
We agree that there are likely to be potential improvements from transformer models. Yet the large sets that are typically required for transformer models are currently not available for whole brain neuroanatomy. A complementary issue is the explainable nature of our approach, which is not achievable with current deep learning methods.
Reviewer 4.
- It is unclear whether each cell image contains only one cell, … .
The reviewer correctly notes that each cell image can contain more than a single neuron. We tried different sized cell images and found that small images lead to neurons that overfill the box. Even if a small fraction of cell images contains multiple neurons, our use of regional features as a training set makes us insensitive to small numbers of errors.
- The feature crafting process could benefit from an ablation study … .
We agree that this is a useful test, as different features are correlated. We have addressed the analogous issue of feature importance (Fig. 6). Note that we are not allowed to perform additional experiments for this submission.
- The definition of a region’s boundary and its impact … are unclear.
This will be clarified in the final submission.
- It is unclear how the classification score reflects the quantity ratio between different structures … .
This too will be clarified.
- With only three brain images available, cross-validation is crucial … .
Our comparison is against the CNN used by [6] and that work did not use cross validation. Again, we are not allowed to perform additional experiments for this submission.
- A comparison of the proposed method’s computational overhead … .
While the performance of our method - in terms of generalization - is greater (Fig. 5), the computational time to build and train the models are comparable.
- Reviewer 3 mentions several specific improvements. We will incorporate them all.
Meta-Review
Meta-review not available, early accepted paper.