LUVS-Net: A Lightweight U-Net Vessel Segmentor for Retinal Vasculature Detection in Fundus Images
, Khuram Naveed 2,3, Ali Nauman 4,†, Sardar Muhammad Gulfam 5 and Sung Won Kim 4,*Round 1
Reviewer 1 Report
In this paper, evaluates the performance of LUVS-Net for retinal vasculature detection in fundus images. It is a good idea, however, there is a lack of theoretical analysis and experimental verification for the proposed framework.
(1) We identified a prior study similar to the authors’ paper. Please explain the difference from the previously published paper.
* Gargari et al., Segmentation of retinal blood vessels using U-Net++ architecture and disease prediction, Electronics, 11, 2022, 3516, https://doi.org/10.3390/electronics11213516
(2) Line 60. Please correct this reference. “That involves development of a pipeline of operations tar-geted at improved detection of eye diseases [14? ]”
(3) Recently, various deep-learning networks have been introduced, so why did you use the existing U-net? If possible, it is recommended to use the recently introduced network to derive results.
* T. Tsai and S. Huang, Refined U-net: a new semantic technique on hand segmentation, Neurocomputing, 495, 2022, 1-10, https://doi.org/10.1016/j.neucom.2022.04.079
* A. Wang, R. Togo, T. Ogawa, M. Haseyama, Defect detection of subway tunnels using advanced U-net network, Sensors, 22, 2022, 2330, https://doi.org/10.3390/s22062330
* Gargari et al., Segmentation of retinal blood vessels using U-Net++ architecture and disease prediction, Electronics, 11, 2022, 3516, https://doi.org/10.3390/electronics11213516
(4) Are there any limitations of the proposed approach? Describe in the discussion session.
(5) Please explain the detail information of used LUVS-Net architecture and parameters (i.e., batch size, learning rate, loss function, optimization solver, etc.)
(6) In order to improve the accuracy of the results, it is recommended to consider additional measurements of the dice coefficient, hauss-dorff distance, mahalanobis distance.
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The LUVSNet is proposed to improve the computational complexity of vanilla U-Net, however, no experiments are provided to verify this. It is necessary to compare LUVSNet against U-Net in terms of, e.g., FLOPs, # model params.
For the baselines in Tables4-7, I am curious why the vanilla U-Net is not compared.
Related work are not comprehensive. Some recent segmentation methods should be discussed in Sec. 2, like Volumetric memory network for interactive medical image segmentation and Rethinking Semantic Segmentation: A Prototype View.
Citation errors in Line 60 should be corrected.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 3 Report
The paper proposes an interesting work on retinal vessel segmentation in fundus images. The authors have developed a novel network that has the advantage of using fewer trainable parameters with which better performance is achieved compared to existing segmentation methods which use a larger number of trainable parameters. The work is somewhat well presented, with experimental evaluations done on three different datasets.
However, there are some concerns:
· It is not clear why the input image is followed by a ReLU layer according to Figure 1, before applying any convolution to the image.
· Grammatical corrections are needed. There are errors on lines 60, 63, 68, 160, 217, 222, 264, and many more.
· The tables are not mentioned in order in the text. Table 1 is mentioned after Table 3. It would be better to number tables in the order they appear in the text and position them accordingly.
· Line 179 says Table 3 shows the LUVS-Net encoder structure, but Table 3 gives the details about the datasets used.
· In line 209, it says “Each layer consists of …”. Isn’t it the architecture that consists of all these?
· Repetition of the words “ground truth” in captions of Figures 2, 3 and 4.
· Table 4 which has been captioned as parameter details of proposed architecture is the same as Table 5. Further, Table 4 is not referenced in the paper.
· A plot showing the test results of the LUVS-NET on the three datasets would be good.
· Bar graphs corresponding to the tables showing comparison with state-of-the-art methods would be good.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
The authors have done revisions to the paper by addressing the comments on the first version. Overall, the paper looks good. There are few spelling errors (“light-weigh” instead of “light-weight” in line 112, “purposed” instead of “proposed” in Table 1 caption and line 228), and few grammatical errors (lines 144, 253, 349) that need to be corrected.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
