Shape Matters: Detecting Vertebral Fractures Using Differentiable Point-Based Shape Decoding

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Detecting vertebral fractures is a crucial aspect of medical imaging and orthopedic research, particularly in the context of conditions such as osteoporosis. This research field is indeed active and valuable.

The proposed shape-based classification framework, incorporating a novel point-based shape decoder, allows deep learning models to be independent of variations in image intensities and reduces the demand for labeled data. They highlight that their approach, which leverages shape information from large-scale multi-label segmentation models, yields preferable results compared to training on image-intensity data directly. The main contributions of the study include an analysis of different auto-encoder architectures, the development of a modular encoder-decoder framework, the introduction of the point-based shape decoder, and extensive experiments to compare different encoder and decoder architectures.

The manuscript is well written. The introduction and discussion sections are quite comprehensive. Overall, this is an important technical paper. I only have a few questions and comments.

 

1. The online source code is incomplete. Please enhance it in the future.
2. The manuscript primarily discusses the proposed methods for computer-aided detection of vertebral compression fractures and evaluating the performance of different encoder-decoder architectures. However, I did not find the definitions for true positive and false negative in the evaluation section. Could you please specify how false positives and true positives are calculated in the context of the study? That will be helpful for readers to understand.
3. The network input to some extent relies on pre-segmentation based on Fig. 1. Will the segmentation accuracy impact the performance of your job? Since the input, S_in, is an initial image for the convolutional neural network, is that possible to use the original image instead of segmentation outcomes to reduce segmentation bias? Will it be better?
4. I noticed your statement: 104 volumes of interest, including 24 vertebrae (C1-L5). From this dataset, we extracted patches, resulting in nearly 13k vertebrae surface masks used for the unsupervised training of our AEs. However, in Table 1, I observed that your model has a few million parameters. Could you provide some discussion on how you train such a large number of parameters with only 13k examples and elaborate on strategies to avoid overfitting?

Author Response

We appreciate the time and effort that you have dedicated to evaluating our work. Below, we provide a detailed response to your comments:

1. Incomplete source code: 
   • We are planning to add our model weights and a demo notebook to our GitHub and add more instructions in the ReadMe.
2. Definition of TP and FP:
   • We explain how we split the labels into binary classifiers in Section 4.2.
3. Impact of segmentation step:
   • This is an important point, which we already addressed by comparing classification results between ground truth masks and segmentation masks using the TotalSegmentator (see Tab.2). 
   • As we point out in the discussion section, the quality of the segmentation does have an impact on the classification task. However, this impact is smaller for convolutional- and point-encoder models compared to graph-based encoder models. 
   • Training an image-based encoder to extract the shapes from image-intensity patches is possible after the localisation of the vertebrae in the scan. The model would need to extract the shape features from the image which complicates the task and is less robust against domain shift. We would also like to point out that segmentation models have an advantage over AE models (skip-connections and more parameters).
4. Overfitting:
   • We would like to clarify that the large majority of trainable parameters in our pipeline are trained on reconstructing the vertebra shape on a dataset consisting of over 13.000 vertebrae. In the second step, we freeze the encoder parameters and train the MLP, which only comprises 541 thousand free parameters, on another dataset.

Reviewer 2 Report

Comments and Suggestions for Authors

Shape Matters: Detecting Vertebral Fractures Using Differentiable Point-Based Shape Decoding

This research emphasizes on the usage of shape auto-encoders with pretraining in unsupervised manner for diagnosis of vertebral Fractures. This work proposes a novel point-based shape decoder to improve detection accuracy and reduce reliance of model on annotated data. It is an interesting research topic as an accurate and early detection of osteoporotic fractures can be beneficial for reducing relevant heath issues.

The paper is well-written, and content is well organized. However, following minor comments need to be addressed:

·       Overall, introduction section is well written and novelty/contributions are presented in a proper way, but the language of the paper requires improvement, as it contains issues such as sentence structure and incorrect spellings. Consider seeking assistance from a native English speaker for proofreading and refinement. Few examples of spelling mistakes or text need to be updated are as follows:

o   Spelling of localization, analyse

o   At line 67, emphasizing

o   Line “The experiments and results Section 4 elaborates”

 

·       Methods are well explained with the detailed comparison of results. However, a comparative study can be added to improve the quality of manuscript by comparing the obtained results with the other research papers.

 

 

 

Comments on the Quality of English Language

The language of the paper requires improvement, as it contains issues such as sentence structure and incorrect spellings. Consider seeking assistance from a native English speaker for proofreading and refinement.

Author Response

We appreciate the time and effort that you have dedicated to evaluating our work. Below, we provide a detailed response to your comments:

1. Spelling: 
   • We would like to point out that we intentionally used British English spelling, which is in adherence to the guidelines provided by the journal.
2. Comparative results with other research papers:
   • We highlighted the PAE - Paper (Sekuboyina et. al. 2019) [12] as the most relevant previous work. The results can be found in the Discussion Section. However, the results are computed on a private dataset and the source code was not published. Furthermore, we would like the point out that the point-encoder point-decoder model is similar to their approach and the reported AUC of their work is not directly comparable.
3. Language:
   • We have checked the manuscript for any inconsistencies regarding British and American English spelling and revised the text for any grammatical issues.

 

[12] Sekuboyina, A.; Rempfler, M.; Valentinitsch, A.; Loeffler, M.; Kirschke, J.S.; Menze, B.H. Probabilistic point cloud reconstructions for vertebral shape analysis. In Proceedings of the Medical Image Computing and Computer Assisted Intervention–MICCAI 2019, Springer

Reviewer 3 Report

Comments and Suggestions for Authors

The article presents research on autoencoder architectures for effective shape encoding of vertebrae to diagnose osteoporotic fractures of vertebrae. The article is well written and demonstrates a concentrated analysis of experimental results. However, the review on related work is uncomprehensive and the experiments lack comparison with methods of other type.  Please find the comments on possible improvements below.

-          (54 line) “recently proposed AE model” – the meaning of AE should be explained with the first occurrence. Similarly, with other abbreviations like CT, MLP, CNN.

-          (2 section) The review of the related work should be more comprehensive (the methods can be grouped by different approaches used, the scope of application and the specific features of the selected problem can be discussed).

-          (4 section) in the Experiments section it would be useful to have methods of other types for detecting vertebral body fractures (e.g. CNN/LSTM, random forest, SVM).

Author Response

We appreciate the time and effort that you have dedicated to evaluating our work. Below, we provide a detailed response to your comments:

1. Abbreviations:
   • We have included your suggested changes in the manuscript
2. Related works:
   • We have subdivided the Related Works Section to improve the structure and readability.
3. Experiments with other methods:
   • We would like to point out that the current SOTA methods employ supervised, image-based CNN models and comparison to other methods (using LSTM etc.), that demonstrate the superiority of recent deep learning approaches against classical machine learning algorithms, which has already been reported in previous work. 

Reviewer 4 Report

Comments and Suggestions for Authors

Thanks for sharing an interesting study. Here are my comments below.

1. Graphical representation Figure 1 and 2 should be the method section.

2. A summary of dataset information has been shown in the earlier part of the method section.

3. Detail of affine augmentation should be provided.

4. Equipment and software information are missing.

5. Comparison of the previous studies, particularly in performance, should be provided in discussion section.

 

 

 

Author Response

We appreciate the time and effort that you have dedicated to evaluating our work. Below, we provide a detailed response to your comments:

1. Figure 1 and 2 in method section:
   • We have moved the figures accordingly.
2. Dataset information in method section:
   • We believe that the provided information about the dataset are important for understanding the method section.
3. Affine Augmentation:
   • We apply an affine transformation on the dataset. The implementation details can be found in our GitHub repository.
4. Equipment and software information:
   • We will include detailed information in the ReadMe of our GitHub repository
5. Comparison to previous studies:
   1. In the discussion section, we highlight the probabilisticAE - Paper (Sekuboyina et. al. 2019) [12] as the most relevant previous work. However, the results are computed on a private dataset and the source code was not published. Furthermore, we would like the point out that the point-encoder point-decoder model is similar to their approach and the reported AUC of their work is not directly comparable.

 

[12] Sekuboyina, A.; Rempfler, M.; Valentinitsch, A.; Loeffler, M.; Kirschke, J.S.; Menze, B.H. Probabilistic point cloud reconstructions for vertebral shape analysis. In Proceedings of the Medical Image Computing and Computer Assisted Intervention–MICCAI 2019, Springer