Optimized Deep Convolutional Neural Networks for Identification of Macular Diseases from Optical Coherence Tomography Images

Round 1

Reviewer 1 Report

This manuscript describes a computational work on the design and validation of pre-trained deep neural networks (DNN) for processing optical coherence tomography (OCT) images. The authors developed a strategy to modify deep neural networks for retinal images using spectral-domain optical coherence tomography. The basic idea is to remove some deep convolutional layers in pre-trained networks, namely GoogLeNet, ResNet and DenseNet, and replace the removed layers with classification layers. The authors describe the architecture, building blocks, parameters and how to make modification for each DNN, as well as discussion of sub-networks. The proposed DNNs are optimized and applied to small-scale and large-scale OCT datasets for the examination of performance. Overall, this is a good piece of work with innovative ideas and solid results. It should be accepted for publication in the Algorithms. However, I do have some questions and suggestions.

Line 15, DNN is defined from the pre-trained deep neural networks. Line 64, DNN is defined from the deep convolutional neural networks. The authors may make sure that the definition of DNN has no confusion.

The authors should give credit of Figure 1. Are those images from their own work? From collaborators? Or from a literature?

Page 7, Line 190-191, “two different OCT datasets”. Since this is the first place the manuscript mentions two OCT datasets, it will be a good idea to specify the large-scale dataset and small-scale dataset.

Although the authors give background information in the Introduction section, it is still unclear why they pick up three DNNs (Inception-v3, ResNet50 and DenseNet121). Is it just because they are modern DNNs? Or there are some other reasons.

Line 25 and line 86: It may be a good idea to put the definition of CNN in line 86, not in the keywords.

Line 113, there should be a space after the period. There are also “space” problems in other locations.

Comments for author File: Comments.pdf

Author Response

Point 1: Line 15, DNN is defined from the pre-trained deep neural networks. Line 64, DNN is defined from the deep convolutional neural networks. The authors may make sure that the definition of DNN has no confusion.

Response 1: DNN is defined from the deep neural networks in Line 15. The redefinition of DNN in Line 65 is deleted.

Point 2: The authors should give credit of Figure 1. Are those images from their own work? From collaborators? Or from a literature?

Response 2: We give credit of Figure 1 in the revised manuscript in Line 46. The images are picked from a public dataset.

Point 3: Page 7, Line 190-191, “two different OCT datasets”. Since this is the first place the manuscript mentions two OCT datasets, it will be a good idea to specify the large-scale dataset and small-scale dataset.

Response 3: We specify the large-scale dataset and small-scale dataset in the Line 198 when the new manuscript first mentions two OCT datasets.

Point 4: Although the authors give background information in the Introduction section, it is still unclear why they pick up three DNNs (Inception-v3, ResNet50 and DenseNet121). Is it just because they are modern DNNs? Or there are some other reasons.

Response 4: In the introduction section(line 101-105), we briefly discuss the reason why we choose Inception-v3, ResNet50 and DenseNet121 in our experiments. It's because they are modern DNNs which balance computational efficiency and classification accuracy.

Point 5: Line 25 and line 86: It may be a good idea to put the definition of CNN in line 86, not in the keywords.

Response 5: We put the definition of CNN in line 88 and remove the definition in the keywords.

Point 6: Line 113, there should be a space after the period. There are also “space” problems in other locations.

Response 6: We check the article and place "space" after the periods.

Reviewer 2 Report

The authors presented a Deep Learning procedure to classify OCT images of the retina among different diseases affecting its structure. The authors employed a fine tuning optimization of pre-trained standard networks which involved also  substitution of the last layers of the machinery. The work is scientifically sound,  well presented and readable. 

The only  concern i have is the absence of any statistical inference on the results.

When the authors claim that one procedure has a higher accuracy of another, i would support it with some statistical tests which imply the measurement of a confidence interval for the accuracy. This could be done for example by means cross-validation procedure.

Author Response

Point 1: When the authors claim that one procedure has a higher accuracy of another, i would support it with some statistical tests which imply the measurement of a confidence interval for the accuracy. This could be done for example by means cross-validation procedure.

Response 1: In the revised version, we support the improvement of sub-networks with one-way analysis of variance with α= 0.05. The details are shown in line 272-281.

Round 2

Reviewer 1 Report

The authors made a revision according to my previous comments. They corrected many "space" problems, but there are still a lot as I list below:

Line 34 before (AMD) and (DME). Line 35 before [5,6]. Line 38before (CNV). Line 43 before [7]. Line 55 before (CAD). Line 59 before (LBP). And before (SP). Line 60 before (PCA) and [9]. Line 63 before (HoG), (SVM) and [12]. Line 65 before (K-SVD) and [13]. Line 67 before [14]. Line68 before [16], [17,18], and [19]. Line 69 before [20]. Line 70 before [21-23]. Line 85 before [24]. Line 87 before [25]. Line 89 before (CNN). Line 99 before [27,28]. Line 114 after DNNs. Line 121 after kernels. Line 125 after convolutions. Line 213 and 214. Line 216, Table 4, after IBDL. Line 256 after images. Line 257 after images. Line 262, Table 5 after CNN and IBDL. Also many are in References.

Since these are just format problems. I am OK to recommend an acceptance of the paper. 

Author Response

Point 1: The authors made a revision according to my previous comments. They corrected many "space" problems, but there are still a lot as I list below:

Line 34 before (AMD) and (DME). Line 35 before [5,6]. Line 38 before (CNV). Line 43 before [7]. Line 55 before (CAD). Line 59 before (LBP). And before (SP). Line 60 before (PCA) and [9]. Line 63 before (HoG), (SVM) and [12]. Line 65 before (K-SVD) and [13]. Line 67 before [14]. Line68 before [16], [17,18], and [19]. Line 69 before [20]. Line 70 before [21-23]. Line 85 before [24]. Line 87 before [25]. Line 89 before (CNN). Line 99 before [27,28]. Line 114 after DNNs. Line 121 after kernels. Line 125 after convolutions. Line 213 and 214. Line 216, Table 4, after IBDL. Line 256 after images. Line 257 after images. Line 262, Table 5 after CNN and IBDL. Also many are in References.

Response 1: We check the "space" problems again. In the revised version, we put "space" after periods and commas in sentences. We also add "space" before open parentheses and open brackets. The References are formatted according to the journal style and "space" are checked as well.

Reviewer 2 Report

The manuscript can be accepted for publication.

In the final version i would report p-values in exponential form and  both p vaules and F rounded to an order of magnitude of 10^-3 (reporting F value up to 10^-8 is a bit too much).

Author Response

Point 1: In the final version i would report p-values in exponential form and  both p vaules and F rounded to an order of magnitude of 10^-3 (reporting F value up to 10^-8 is a bit too much).

Response 1: We report p-values in exponential form in the revised version (line 274-275). F and p values are rounded to an order of magnitude of 10^-4 and 10^-3 respectively.