DLBCNet: A Deep Learning Network for Classifying Blood Cells
Round 1
Reviewer 1 Report
This paper is very hard to follow. Authors are requested to modify the paper as follows:
1. ResNet50 is used as the backbone. The author should explain why and how it is done.
2. In BCGAN, The generator and discriminator are used. the author has to explain how these are created. what are the reasons to select those layers?
3. The last five layers of the backbone model are replaced in the Proposed ETRN. what are the problems of those five layers? how did you recover with ETRN?
4. The Proposed DLBCNet is not clear. please clearly explain it. the author must explain it from a technical point of view.
5. The proposed DLBCNet is compared to other state-of-the-art methods [17-20]. Two of those are conference proceedings. it would be better to compare the proposed work with standard publications (like BDCC)
Author Response
- ResNet50 is used as the backbone. The author should explain why and how it is done.
Response: Thank you for your suggestion. We write the reason in the Section 4.3.
‘The ResNet50 as the backbone model could get the best results than other back-bone models. The reason is that the residual connection in the ResNet50 can improve the classification performance. Even though the residual connection is still in the Res-Net18, deeper layers could extract better features. In this situation, the ResNet50 as the backbone model has better performance than the ResNet18.’
- In BCGAN, The generator and discriminator are used. the author has to explain how these are created. what are the reasons to select those layers?
Response: Thank you for your comment. We explained it in the Section 3.2.
‘In the proposed BCGAN, the convolutional layers are used to extract features. The LeakyReLu is implemented to add the nonlinearity. Max pooling is a common strategy to downsample the extracted features. Batch normalization (BN) is chosen to alleviate the gradient disappearance. The overfitting problem could be alleviated by adding the dropout. The BCGAN is specially designed for blood cell images.’
- The last five layers of the backbone model are replaced in the Proposed ETRN. what are the problems of those five layers? how did you recover with ETRN?
Response: Thank you for your advice. We explained it in Section 3.3.
‘These three RNNs merely include three layers, which are the input layer, hidden layer, and output layer. The training of RNNs could be faster than traditional CNN models benefiting from the simple structure. Compared with the back-propagation neural network, because the weights and bias in RNNs have been randomly initialized and fixed in training and the outputs can be calculated by pseudo-inverse, it is unnecessary to update the parameters based on back-propagation, which could shorten the training time. On the other hand, these three RNNs used to replace the last five layers can im-prove the classification performance.’
- The Proposed DLBCNet is not clear. please clearly explain it. the author must explain it from a technical point of view.
Response: Thank you for your suggestion. We explain it in the Section 3.4.
‘We propose a novel DL network to diagnose blood cells (DLBCNet). Collecting a large number of labeled blood cell images to train DL modes is a challenge due to the cost and time restrictions. We propose a new specifical model for blood cells (BCGAN) to cope with this challenge. More filters and dropout layers for each layer are added to capture more high-level features. Additional dropout layers and BN are added to avoid the overfitting problem.
Meanwhile, the checkboard patterns could be alleviated by the biggest kernel size. The ResNet50 pre-trained on the ImageNet dataset is implemented as the backbone model in this paper, which is modified and fine-tuned based on blood cells because of the difference of the ImageNet dataset with the blood cell dataset used in this paper. The modified ResNet50 is applied as the feature extractor. The last five layers of the modified ResNet50 is substitute with three RNNs (ELM, RVFL, and SNN). These three RNNs are used for classification. The sample structure and randomized weights of RNNs could reduce training time.
Nevertheless, the RNN is considered an unstable neural network due to some randomized operations. We propose the ETRN by combining three RNNs based on the majority voting to improve the robustness and the generalization performance.’
- The proposed DLBCNet is compared to other state-of-the-art methods [17-20]. Two of those are conference proceedings. it would be better to compare the proposed work with standard publications (like BDCC)
Response: Thank you for your suggestion. These two conferences are deleted. More standard publications like BDCC are discussed in Introduction.
Reviewer 2 Report
The title “ DLBCNet: A Deep Learning Network for Classifying Blood 2 Cells” is quite interesting but I have some concerns on this work which are given below.
1: Please make proper space between abstract and keywords.
2: Please re-write or improve the 1st paragraph of the introduction. Please mention the role of haemoglobin briefly. The 2nd paragraph is also not up-to the mark.
3: Please correct the English mistakes properly e.g. Suitable prepositions.
4: In introduction, in the limitation part, you said DL is not good and have higher chances of overfitting but you have also used the same DL technique. Please justify here briefly.
5: Why the figure 1 sections (a-d) are not on the same pattern.
6: I’ll suggest you to improve the figure 5.
7: It will be better if you add the area under the curve of the models.
8: Please provide the training code of the model via online repositories like GitHub or Zenodo.
Author Response
1. Please make proper space between abstract and keywords.
Response: Thank you for your comments. We just make proper space between abstract and keywords.
2. Please re-write or improve the 1st paragraph of the introduction. Please mention the role of haemoglobin briefly. The 2nd paragraph is also not up-to the mark.
Response: Thank you for your comments. We re-write the first and second paragraph in the Introduction.
‘The blood flowing in blood vessels is combined with blood cells and plasma. Blood is red because of red blood cells in the blood. Hemoglobin is a special protein that transports oxygen within red blood cells. It is a protein that makes the blood red and consists of globin and heme. Besides red blood cells, there are also white blood cells and platelets. Although they occupy a small share, their functions are very important. These three kinds of blood cells account for 45% of the blood volume, and the remaining 55% of the volume is plasma.
Blood is distributed throughout the body and delivers nutrients to various organs. Naturally, it also stores important health information about the human body. The blood composition will change when there is a problem in our body. Therefore, the diagnosis of blood can indirectly help doctors judge a person's physical state, which is the routine blood test we often hear. The routine blood test mainly includes the diagnosis of red blood cells, white blood cells, and so on. Its significance is to find many early signs of systemic diseases, diagnose whether there is anemia or blood system dis-ease, and reflect the hematopoietic function of bone marrow. Mainstream blood diagnosis is now used to detect white blood cell abnormalities. White blood cell analysis is an essential examination method for pathological blood samples and is an important indicator for detecting and observing diseases. White blood cell recognition is one of the important contents of blood testing. By identifying the total number, relative ratio, and morphology of various white blood cells in the blood, we can determine whether there is a disease, the type of disease, and the severity of the disease. So, the examination of white blood cells is very important to understand the body's condition and diagnose diseases.’
3. Please correct the English mistakes properly e.g. Suitable prepositions.
Response: Thank you for your suggestion. We try our best to correct the English mistakes.
4. In introduction, in the limitation part, you said DL is not good and have higher chances of overfitting but you have also used the same DL technique. Please justify here briefly.
Response: Thank you for your advice. We justify it in the introduction.
‘This paper demonstrates a novel DL model (DLBCNet) for the multi-classification of blood cells. We use pre-trained ResNet50 as the backbone to extract ideal features. There are two ways to deal with the overfitting problem in this paper. First, we propose a new model (BCGAN) to generate synthetic images to create a larger dataset. Second, the proposed ETRN not only has a simpler structure but also achieves better performance than common DL models.’
5. Why the figure 1 sections (a-d) are not on the same pattern.
Response: Thank you for your comment. We revise it and keep them consistence.
6. I’ll suggest you to improve the figure 5.
Response: Thank you for your suggestion. We redraw the figure 5.
7. It will be better if you add the area under the curve of the models.
Response: Thank you for your suggestion. We added the ROC and AUC in the Section 4.2.
‘Five multi-classification measurements are implemented to evaluate the proposed DLBCNet. Considering the contingency, we carry out five runs. The classification performance of the proposed DLBCNet by five runs is presented in Table 7. The average accuracy, sensitivity, precision, specificity, and f1-score per class by five runs are given in Table 8. The average-accuracy, average-sensitivity, average-precision, average-specificity, and average-f1-score of the proposed model are 95.05%, 93.25%, 97.75%, 93.72%, and 95.38%, accordingly. All the measurements per class of the proposed DLBCNet are greater than 90%. Especially, our model can achieve the promising average-accuracy for each class. The ROC curve is presented in Figure 7. The AUC values for eosinophil, lymphocyte, monocyte, and neutrophil are 0.8922, 0.9957, 0.9694, and 0.9091.’
8. Please provide the training code of the model via online repositories like GitHub or Zenodo.
Response: Thank you for your suggestion. The link is Zhu-tubage/DLBCNet (github.com).
Reviewer 3 Report
The paper proposes a deep learning network to classify blood cells. Experimental results demonstrate the effectiveness of the proposed method. Overall, the paper is well written. Some concerns are as follows.
1. The part of contribution should be written to demonstrate what problems the proposed model have addressed rather than what they do.
2. Figures in the manuscript should be enlarged to make it readable.
Author Response
- The part of contribution should be written to demonstrate what problems the proposed model have addressed rather than what they do.
Response: Thank you for your suggestion. We re-write the contributions in the Introduction.
‘This paper demonstrates a novel DL model (DLBCNet) for the multi-classification of blood cells. We use pre-trained ResNet50 as the backbone to extract ideal features. There are two ways to deal with the overfitting problem in this paper. First, we pro-pose a new model (BCGAN) to generate synthetic images to create a larger dataset. Second, the proposed ETRN not only has a simpler structure but also achieves better performance than common DL models. The main contributions of our jobs are given as follows:
- The pre-trained ResNet50 is implemented to extract ideal features by comparing it with other CNN models.
- The proposed BCGAN is used to generate synthetic images to alleviate the over-fitting problem.
- We propose ETRN to enhance the robustness with the ensemble strategy of com-bining three individual networks.
- We propose a novel DL model to classify blood cells, which is named DLBCNet.’
- Figures in the manuscript should be enlarged to make it readable.
Response: Thank you for your suggestion. We try our best to improve the figures.
Round 2
Reviewer 1 Report
In response to question No. 3, you mention ' Compared with the back-propagation neural network, because the weights and bias in RNNs have been randomly initialized and fixed in training and the outputs can be calculated by pseudo-inverse, it is unnecessary to update the parameters based on back-propagation, which could shorten the training time.'
This problem is very common and still exists in your backbone network (except the last 5 layers). Explain what were the problems of the last five layers.
Author Response
- This problem is very common and still exists in your backbone network (except the last 5 layers). Explain what were the problems of the last five layers.
Response: Thank you for your comments. We answer this comment in Section 3.3 and 4.5.
‘The training of RNNs could be faster than traditional CNN models benefiting from the simple structure. Compared with the back-propagation neural network, because the weights and bias in RNNs have been randomly initialized and fixed in training and the outputs can be calculated by pseudo-inverse, it is unnecessary to update the parame-ters based on back-propagation, which could shorten the training time. On the other hand, these three RNNs used to replace the last five layers can improve the classifica-tion performance.’
‘Three RNNs are implemented as the classifier to replace the last five layers of the backbone model, which are ELM, RVFL, and SNN. The training time of RNNs could be less than traditional CNN models because of the simple structure and fixed random-ized parameters. At the same time, RNNs could achieve promising results.
The effects of RNNs are given in Table 11. The classification results using the last five layers are not as good as those using three RNNs. It is very clear to conclude that three RNNs used to substitute the last five layers can achieve better classification per-formance. The RNNs can have positive effects on blood cell classification.’
