Lung Nodule CT Image Segmentation Model Based on Multiscale Dense Residual Neural Network
Round 1
Reviewer 1 Report
In this paper, the authors propose a custom UNET model to segment the lung nodule from CT images. The paper performs some incremental architecture changes compared to the standard UNET. Several technical and methodological issues with the paper must be addressed before the resubmission. The concerns are presented below:
- Instead of dividing the images into parts, why don't the authors directly process the input images? For robust segmentation, the authors need to consider the whole input image rather than processing them into parts. Also, Maybe the additional accuracy reported in the paper is a result of the preprocessing stage, which reduces the complexity of the image being processed.
- Please be advised that for an accurate performance evaluation, the authors must perform a k-fold cross-validation technique, as one set evaluation may bias the results. Also, it is not clear what is the validation set used for, as the authors don't perform any hyperparameter tuning. It is reported that the authors train the model for 240 epochs. How do the authors dictate if this is an ideal epoch for model convergence? How do they ensure that the model is not underfitting/overfitting?
- The datasets comprise three partitions in a standard ML pipeline: train, test, and valid. The model is trained with a training set and performed hyperparameter tuning/early stopping with a validation set. Once the optimal model configuration is estimated, the model is tested with an unseen test partition. The train-test-valid folds can also be derived with k-fold sampling. The authors don't practice this standard in the paper, and the reviewer suspects that the results reported on the paper are biased for a specific training test set. The authors need to report the k-fold performance in the paper.
- Do the authors evaluate the rest of the benchmark frameworks on the processed images (chunked images) or the original unprocessed image?
- The authors also need to present some of the visual segmentation output results of the proposed framework and contrast the inaccuracies in segmentation with the proposed framework in comparison to others.
Author Response
Point 1: Instead of dividing the images into parts, why don't the authors directly process the input images? For robust segmentation, the authors need to consider the whole input image rather than processing them into parts. Also, Maybe the additional accuracy reported in the paper is a result of the preprocessing stage, which reduces the complexity of the image being processed.
Response 1: Since the images in the LIDC-IDRI data set are lung CT images, one image contains a lot of information, and there may be multiple lung nodules of different sizes. Without data preprocessing, the features of lung nodules cannot be extracted separately. Based on the suggestions of reviewers, the paper redesigned the operation of image preprocessing and reached the result that after data processing, one image only contained one pulmonary nodule.The new draft is uploaded as an attachment.
Point 2: Please be advised that for an accurate performance evaluation, the authors must perform a k-fold cross-validation technique, as one set evaluation may bias the results. Also, it is not clear what is the validation set used for, as the authors don't perform any hyperparameter tuning. It is reported that the authors train the model for 240 epochs. How do the authors dictate if this is an ideal epoch for model convergence? How do they ensure that the model is not underfitting/overfitting?
Response 2: Based on the suggestions of reviewers, 10-fold cross-validation technology has been added, and the basis for hyper-parameter tuning and judging the convergence of the model has been stated in the attachment.
Point 3: The datasets comprise three partitions in a standard ML pipeline: train, test, and valid. The model is trained with a training set and performed hyperparameter tuning/early stopping with a validation set. Once the optimal model configuration is estimated, the model is tested with an unseen test partition. The train-test-valid folds can also be derived with k-fold sampling. The authors don't practice this standard in the paper, and the reviewer suspects that the results reported on the paper are biased for a specific training test set. The authors need to report the k-fold performance in the paper.
Response 3: The k-fold cross-validation technique has been added to the attachment.
Point 4: Do the authors evaluate the rest of the benchmark frameworks on the processed images (chunked images) or the original unprocessed image?
Response 4: All the comparison experiments used preprocessed images
Point 5: The authors also need to present some of the visual segmentation output results of the proposed framework and contrast the inaccuracies in segmentation with the proposed framework in comparison to others.
Response 5: Image segmentation visualization results have been added to the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
In this paper the authors present a lung nodule CT image segmentation method showing good results over a public dataset.
"Cavity" is only mentioned in the abstract, without further reference in the following text.
Line 187: please explain what pylidc is.
Please add references to some of the software you have used: Keras, Tensorflow...
Please discuss the methods presented in
Suji, R.J., Bhadouria, S.S., Dhar, J. et al. Optical Flow Methods for Lung Nodule Segmentation on LIDC-IDRI Images. J Digit Imaging 33, 1306–1324 (2020). https://doi.org/10.1007/s10278-020-00346-w
In lines 181-189 the data augmentation process is presented, but it is not clear if this has been performed before or after the training-test partition of the data set. If the training-test partition has been performed after the data augmentation process, the results are incorrect and misleading, as augmented (and therefore similar) versions of data in the training set could appear in the test set. In my opinion, the data must be partitioned between training and testing, the training augmented, the model built, and then tested against the non-augmented test data.
Author Response
Point 1: "Cavity" is only mentioned in the abstract, without further reference in the following text.
Response 1: The word "Cavity" comes from "jagged cavity convolution", and the section is written as "jagged expansion convolution", which is the same in Chinese. I'm sorry for the inconsistency caused by my negligence. Now the new draft has been uploaded in the attachment and the version has been unified
Point 2: Line 187: please explain what pylidc is.
Response 2: Pylidc is a python toolkit that can be used to analyze and query annotation data (without access to DICOM image data), as well as to manipulate DICOM files associated with LIDC data sets. This process is the only way to read LIDC-IDRI. Considering that it does not have any innovative meaning, it has been deleted from the article.
Point 3: Please add references to some of the software you have used: Keras, Tensorflow...
Response 3:
- TensorFlow is a deep learning framework released by Google. It has received a lot of attention since its release and has been widely used in computer vision, audio processing, recommendation systems, and natural language processing
- Anaconda and TensorFlow Anaconda is an open-source Python distribution that includes many science packages.
- Keras is a high-level Python neural network framework, and its documentation is detailed. Keras has been added to TensorFlow as its default framework, providing a more advanced API for TensorFlow.
- Pylidc is a python toolkit that can be used to analyze and query annotation data (without access to DICOM image data), as well as to manipulate DICOM files associated with LIDC data sets.
Point 4: Please discuss the methods presented in Suji, R.J., Bhadouria, S.S., Dhar, J. et al. Optical Flow Methods for Lung Nodule Segmentation on LIDC-IDRI Images. J Digit Imaging 33, 1306–1324 (2020). https://doi.org/10.1007/s10278-020-00346-w
Response 4: The optical flow method is actually a method to infer the moving speed and direction of objects by detecting the intensity of image pixels over time. It is very useful in pattern recognition, computer vision, and other image processing fields, and can be used for motion detection, object cutting, collision time and object expansion calculation, and motion compensation coding. Or through the object’s surface and edge for three-dimensional measurement and so on. Suji, R.J. et al. proposed that the idea of segmenting pulmonary nodules by optical flow method was novel and could achieve the purpose of segmentation with certain effects.
The model in this paper is semantic segmentation based on morphology. It improves the segmentation accuracy by changing the network structure, the number of network layers, the depth of the network, adjusting parameters, and so on, which is very classic in the processing of medical images.
Point 5: In lines 181-189 the data augmentation process is presented, but it is not clear if this has been performed before or after the training-test partition of the data set. If the training-test partition has been performed after the data augmentation process, the results are incorrect and misleading, as augmented (and therefore similar) versions of data in the training set could appear in the test set. In my opinion, the data must be partitioned between training and testing, the training augmented, the model built, and then tested against the non-augmented test data.
Response 5: Data enhancement occurs after dividing the train-test partition of the data set.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The new text version has several grammatical and sentence structure issues in the paper's body and the image captions. Please carefully review the manuscript and address these issues.
Author Response
point 1 ：The new text version has several grammatical and sentence structure issues in the paper's body and the image captions. Please carefully review the manuscript and address these issues.
answer: Thank you for your comments. Native English speakers have been found to revise the manuscript.
Reviewer 2 Report
The English usage and grammar have to be thoroughly checked. For example, in lines 165-166 "The situation of the LIDC-IDRI dataset is shown..." should be corrected.
Please cite the LIDC-IDRI lung nodule dataset paper:
Armato III, Samuel G., Geoffrey McLennan, Luc Bidaut, Michael F. McNitt‐Gray, Charles R. Meyer, Anthony P. Reeves, Binsheng Zhao et al. "The lung image database consortium (LIDC) and image database resource initiative (IDRI): a completed reference database of lung nodules on CT scans." Medical physics 38, no. 2 (2011): 915-931.
You are working on semantic segmentation but the reader should be aware what are the pros and cons of using semantic segmentation instead of other approaches. Therefore, please expand the introduction or add a new state-of-the-art section where you should cite and discuss previous reviews, as the following paper:
Pehrson, Lea Marie, Michael Bachmann Nielsen, and Carsten Ammitzbøl Lauridsen. "Automatic pulmonary nodule detection applying deep learning or machine learning algorithms to the LIDC-IDRI database: a systematic review." Diagnostics 9, no. 1 (2019): 29.
and also other approaches using other methods, such as optical flow:
Suji, R.J., Bhadouria, S.S., Dhar, J. et al. Optical Flow Methods for Lung Nodule Segmentation on LIDC-IDRI Images. J Digit Imaging 33, 1306–1324 (2020). https://doi.org/10.1007/s10278-020-00346-w
In lines 189-191 does not make sense to name python 3.9 and delete the reference to keras and tensorflow. Please add them back in the main text and refer to them in the bibliography or remove that paragraph. In short, if you write "This experiment is implemented on the Windows 10 operating system. The environment of the experiment is python3.9, the processor is i5-8265U2.30G HzCPU, and the memory is 8GB.", expand it or delete it.
Author Response
- The English usage and grammar have to be thoroughly checked. For example, in lines 165-166 "The situation of the LIDC-IDRI dataset is shown..." should be corrected.
It has been changed to: The LIDC-IDRI dataset is shown in (Table 1).
- Please cite the LIDC-IDRI lung nodule dataset paper:
Armato III, Samuel G., Geoffrey McLennan, Luc Bidaut, Michael F. McNitt‐Gray, Charles R. Meyer, Anthony P. Reeves, Binsheng Zhao et al. "The lung image database consortium (LIDC) and image database resource initiative (IDRI): a completed reference database of lung nodules on CT scans." Medical physics 38, no. 2 (2011): 915-931.
A reference to the LIDC-IDRI dataset paper has been added on lines 245-246.
- You are working on semantic segmentation but the reader should be aware what are the pros and cons of using semantic segmentation instead of other approaches. Therefore, please expand the introduction or add a new state-of-the-art section where you should cite and discuss previous reviews, as the following paper:
Pehrson, Lea Marie, Michael Bachmann Nielsen, and Carsten Ammitzbøl Lauridsen. "Automatic pulmonary nodule detection applying deep learning or machine learning algorithms to the LIDC-IDRI database: a systematic review." Diagnostics 9, no. 1 (2019): 29.
and also other approaches using other methods, such as optical flow:
Suji, R.J., Bhadouria, S.S., Dhar, J. et al. Optical Flow Methods for Lung Nodule Segmentation on LIDC-IDRI Images. J Digit Imaging 33, 1306–1324 (2020). https://doi.org/10.1007/s10278-020-00346-w
In lines 34-144, a general overview of the methods for pulmonary nodule segmentation is added, along with a summary of the advantages and disadvantages of these methods.
- In lines 189-191 does not make sense to name python 3.9 and delete the reference to keras and tensorflow. Please add them back in the main text and refer to them in the bibliography or remove that paragraph. In short, if you write "This experiment is implemented on the Windows 10 operating system. The environment of the experiment is python3.9, the processor is i5-8265U2.30G HzCPU, and the memory is 8GB.", expand it or delete it.
The previous keras and tensorflow sections have been added back into the text and the relevant references have been added.
Author Response File: Author Response.docx
Round 3
Reviewer 2 Report
I believe my suggestions have been addressed. Anyway, I could not find the reference [32] (Tensorflow), so please check that the references are correct.
