Efficient Lung Cancer Image Classification and Segmentation Algorithm Based on an Improved Swin Transformer

Round 1

Reviewer 1 Report

In this paper, the authors proposed a segmentation method based on efficient transformer and applied it to medical image analysis. The algorithm completed the mission of lung cancer classification and segmentation by analyzing lung cancer data, and aimed to provide efficient technical support for medical staff. In addition, the authors evaluated and compared the results in various aspects. For the classification mission, the max accuracy of Swin-T by regular training and Swin-B in two resolutions by pre-training can be up to 82.3%. For the segmentation mission, pre-training is used to help the model improve the accuracy of the experiments. The mIoU of the three models reaches over 45. The experiments demonstrate that the algorithm can be well applied to lung cancer classification and segmentation. 

There are some suggestions to be incorporated in the manuscript

1. Analyze the improvement in the evaluation parameters in comparison of existing techniques

2. Rewrite the conclusion by adding some data of improvement in comparison of the existing techniques applicable to the medical images.

3. Include the parameters tuning used in the experiment.

Author Response

1 Analyze the improvement in the evaluation parameters in comparison of existing techniques.

Reply：In the classification task, we used two sets of ViT models as comparative experiments; in the segmentation task, we used two different backbones, ResNet-101 and DeiT-S, as comparative experiments. By comparing the experiments for the classification task and the segmentation task, we can convincingly demonstrate that Swin Transformer achieves good results in lung cancer detection.

2 Rewrite the conclusion by adding some data of improvement in comparison of the existing techniques applicable to the medical images.

Reply：We have rewritten the conclusions in the abstract by comparing the results of the comparative experiments in lung cancer classification and segmentation with our methodology and marked yellow.

3  Include the parameters tuning used in the experiment.

Reply：The parameters in the chart have been marked to make the data more visible.

Reviewer 2 Report

Efficient Lung Cancer Image Classification and Segmentation 1 Algorithm Based on Improved Swin Transformer 

by

Ruina Sun Yuexin Pang Wenfa Li Hanyu Zhao

The paper discuss about a novel algorithm for lung cancer image classification and segmentation. The method is based on artificial intelligence techniques, more specifically it exploits swin tranformers architectures. The paper is well written, method is clearly described and result are impressive (compared to state-of-the-art methods.)

I have just minor comments:

- Introduction, page 2 line 64: define explicitly the acronym "ViT"

- Section 3.1, page 7, lines 252-256: the statement sound redundant, please rewrite

-Section 3.1, page 7, lines 267-268: I suppose that the "difference" mentioned refers to the population of the two classes. Please write it explicitly.

- Section 3.2, page 8, line 297: what does "top" stands for? You consider top-1 acc and top-5 acc...

- Fig. 4: there is a graphical problem with the sub-labels

- Section 3.3.1, page 9, lines 334-337, the definition of top-1 acc and top-5 acc is unclear.

- Section 3.3.1, page 10, lines 343-346: Fig. 5 shows losses and accuracies of two swin_b architectures. In the text you erroneously mentioned swin-t...

- Section 3.3.1, page 11, line 366: what does it means:" the results are not unsatisfactory"?

Author Response

1 Introduction, page 2 line 64: define explicitly the acronym "ViT"

Reply：I have clearly marked the full name of ViT in brackets.

2 Section 3.1, page 7, lines 252-256: the statement sound redundant, please rewrite

Reply：The passage has been rewritten and marked yellow.

3 Section 3.1, page 7, lines 267-268: I suppose that the "difference" mentioned refers to the population of the two classes. Please write it explicitly.

Reply：In lines 267-268, we have write the “difference” explicitly and marked yellow.

4 Section 3.2, page 8, line 297: what does "top" stands for? You consider top-1 acc and top-5 acc...

Reply：The term "top" refers to the rank of a class in terms of its predicted probability of being the true label. I have further explained top-1 acc and top-5 acc and marked them in yellow.

5 Fig. 4: there is a graphical problem with the sub-labels

Reply：I have corrected the drawing problem in Figure 4 and found a clerical error in the explanatory section of Figure 4 which has been corrected and highlighted in yellow.

6 Section 3.3.1, page 9, lines 334-337, the definition of top-1 acc and top-5 acc is unclear.

Reply：I have detailed the definitions of top-1 acc and top-5 acc at the locations indicated in question 4 and on lines 334-337, and marked yellow.

7 Section 3.3.1, page 10, lines 343-346: Fig. 5 shows losses and accuracies of two swin_b architectures. In the text you erroneously mentioned swin-t...

Reply：I apologize for this clerical error in the text, which has been corrected and marked yellow.

8 Section 3.3.1, page 11, line 366: what does it means:" the results are not unsatisfactory"?

Reply：I have clarified the specific reference to the result and changed the description to ensure that the conclusion is clear.