Methodological Research on Image Registration Based on Human Brain Tissue In Vivo
Round 1
Reviewer 1 Report
Please see the attachment.
Comments for author File: 
Comments.pdf
Author Response
Response to Reviewer 1 Comments
Major comments
Point 1: Methods: I felt Sections are confusing. How about summarizing such as the algorithm description in 2.1-, the analysis in 2.2-, and the database in 2.3- ? In addition, In P.7 L.205, the name of the section (Results) should be changed to avoid confusion with the section containing the experimental results. For example, ”Evaluation index for the registration accuracy” etc.
Response 1：We thank the reviewer for the suggestions. According to the reviewer’s suggestions, we have revised the title of Materials and Methods such as the algorithm description in 2.1, the evaluation index for the registration accuracy in 2.2, and the data source and preprocessing in 2.3. In L.231 on P.7, we have corrected the inappropriate statement. In the revised manuscript, “Results” was changed to “Evaluation index for the registration accuracy”.
Point 2: Methods: Please provide the hardware specs and software type used for this processing.
Response 2：We thank the reviewer for the suggestion. we have added “The experimental environment of this work is configured as: Windows10 operating system, Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz processor and 8GB memory. The realization and test of this study are based on the MatlabR2019a.” in the revised manuscript.
Point 3: Methods: Please compare the computational complexity with different algorithms.
Response 3：We agree with the comment made by the reviewer. According to the steps of our technique, the proposed method is of the same time complexity as the surf algorithm. The surf algorithm is more effective than the sift algorithm and Demons, the details are as follows. Demons does not need to detect feature point, thus avoiding the registration error caused by the mismatch of feature points. However, it depends on all gray values within the images. Therefore, the time complexity of Demons is higher than Sift and Surf. Sift and Surf can perform image registration by a series of feature points, and they thus can effectively discard the redundant information within brain images. The Sift descriptor is a 128-dimensional vector, whereas Surf descriptor is a 64-dimensional vector. Therefore, Sift is slower than Surf for the matching of feature points due to its higher dimensionality. According to the suggestions of the reviewer, we have added “According to the steps of our technique, the proposed method is of the same time complexity as the surf algorithm.” in the revised manuscript.
Point 4: Discussion: Good results would be got with T1 images, is there a clear rationale for using T1 images? If there is, please show it along with error factors such as the image contrast. Also, could you provide some explanations for the versatility of your proposed algorithm? (For T2 and CT images)
Response 4：We thank the reviewer for highlighting the problem. First of all, for the brain MRI, T1 images are more suitable for observing anatomical structures, and T2 images are better for showing tissue lesions. Brain image registration can be divided into two categories according to the target of registration: registration for different individuals and registration for multiple brain images of the same person. The former is to explore a series of transformation parameters for the spatial consistence of the brain shape, size, tissue location and anatomical structure from different individuals. Given the perspective, T1 images can better achieve the goal of spatial consistence than T2 images. Compared to brain MRI, the data acquisition of the brain CT has radiation, which may hurt the human brain. As the reviewer considered, the proposed method can also be applied to T2, CT and other brain images. However, the registration effect needs to be verified. According to the reviewer's comments, we have added the relevant contents in the limitation section.
Point 5: Conclusion: Please tell us about your future work plans.
Response 5: In the future, we will further improve the method, and also try to apply the proposed method to other brain images (such as T2 and CT) for higher universality. We have added the relevant contents in the revised manuscript.
Minor comment
Point 1: Table 1,2: For clarity, please be bold the best values.
Response 1: We thank the reviewer for the suggestion. we have bolded the best values in Table 1 and Table 2 in the revised manuscript as follows.
Author Response File: Author Response.pdf
Reviewer 2 Report
The authors have proposed an image registration workflow for brain tissue on T1 images. Experiments were conducted from publicly available CUMC12 and MGH10 datasets. Results were compared with SIFT, SURF, and Demons algorithms. The proposed method shows improved overall performance than the other three methods in terms of MSD, NCC, NMI, and MI metrics.
Some comments:
1. The introduction part includes informative background. However, there’s not enough literature review of current and state-of-the-art approaches regarding the image registration topic.
2. The baseline methods, SIFT, SURF, and Demons, are relatively traditional. It would be interesting to see a comparison with recent methods.
3. The proposed workflow relies on the segmentation of gray matter and white matter. And Naïve Bayes method was used in the process. Sometimes, Naïve Bayes doesn’t perform very well in several scenarios. Will other segmentation methods, such as U-Net architecture-based methods, help to improve the final performance?
4. The authors claimed the proposed workflow is the closest to the reference image in Figs. 6-9 (lines 290-296). However, it’s visually not easy to tell the difference between the proposed method and SURF, particularly in Figs. 6-8.
5. The authors claimed that the proposed method can “maximally” assure the correspondence of different individual brain images for spatial localization and anatomical structures (lines 348-349). It will be helpful to include more evidence to support the “maximal” argument.
Author Response
Response to Reviewer 2 Comments
Major comments
Point 1:The introduction part includes informative background. However, there’s not enough literature review of current and state-of-the-art approaches regarding the image registration topic.
Response 1: We thank the reviewer for pointing out the problem. According to the reviewer’s suggestion, we have added the related contents in the re-submission as follows.
“By the emerging of deep learning-based approaches, they have changed the landscape of medical image processing research and achieved the-state-of-art performances in many applications(1-2) including brain image registration. Recently, Fan J et al.(3) proposed a deep learning approach for image registration by predicting deformation from image appearance on brain image. Experiments on a variety of datasets showed promising registration accuracy and efficiency. Ma Y et al.(4) proposed an unsupervised deformable image registration network (UDIR-Net) for 3D medical images. Their experiments on brain MR images showed that UDIR-Net exhibited competitive performance against several methods (such as DIRNet, FlowNet, JRS-Net, Elastix and VoxelMorph). Chen J et al.(5) presented ViT-V-Net, which bridges ViT and ConvNet to provide volumetric medical image registration. In brain MRI images, their proposed architecture achieved superior performance to several top-performing registration methods. Song L et al.(6) proposed a hybrid network which combine Transformer and CNN , and applied it to brain MRI images. They found that the method improved the accuracy by 1% compared with the VoxelMorph, ViT-V-Net and SYMNet. The above studies have provided excellent effect on brain registration. However, deep leaning-based registration generally has relatively higher requirements for equipment, environment and data, and its results may vary with one-training iteration. Therefore, given the above, we improved the brain registration based on the traditional method rather than the deep learning method.”
- Fu, Yabo, Yang Lei, Tonghe Wang, Walter J Curran, Tian Liu, and Xiaofeng Yang. "Deep Learning in Medical Image Registration: A Review." Physics in Medicine & Biology 65, no. 20 (2020): 20TR01.
- Abbasi, Samaneh, Meysam Tavakoli, Hamid Reza Boveiri, Mohammad Amin Mosleh Shirazi, Raouf Khayami, Hedieh Khorasani, Reza Javidan, and Alireza Mehdizadeh. "Medical Image Registration Using Unsupervised Deep Neural Network: A Scoping Literature Review." Biomedical Signal Processing and Control 73 (2022): 103444.
- Fan, Jingfan, Xiaohuan Cao, Pew-Thian Yap, and Dinggang Shen. "Birnet: Brain Image Registration Using Dual-Supervised Fully Convolutional Networks." Medical image analysis 54 (2019): 193-206.
- Mahapatra, Dwarikanath, and Zongyuan Ge. "Training Data Independent Image Registration Using Generative Adversarial Networks and Domain Adaptation." Pattern Recognition 100 (2020): 107109.
- Chen, Junyu, Yufan He, Eric C Frey, Ye Li, and Yong Du. "Vit-V-Net: Vision Transformer for Unsupervised Volumetric Medical Image Registration." arXiv preprint arXiv:2104.06468 (2021).
- Song, Lei, Guixia Liu, and Mingrui Ma. "Td-Net: Unsupervised Medical Image Registration Network Based on Transformer and Cnn." Applied Intelligence (2022): 1-9.
Point 2: The baseline methods, SIFT, SURF, and Demons, are relatively traditional. It would be interesting to see a comparison with recent methods.
Response 2: We thank the reviewer for the comment. As the reviewer described, it would be interesting to observe a comparison with the recent methods (deep learning-based method). By the emerging of deep learning-based approaches, they have changed the landscape of medical image processing research and achieved the-state-of-art performances in many applications including brain image registration. Compared with traditional registration methods, deep leaning-based registration has great advantages and potential(1-2) . However, deep leaning-based registration generally has relatively higher requirements for equipment, environment and data, and its results may vary with one-training iteration. Therefore, given the above, we improved the brain registration based on the traditional method rather than the deep learning method. According to the suggestions of the reviewer, we have added relevant contents in our re-submission.
- Fu, Yabo, Yang Lei, Tonghe Wang, Walter J Curran, Tian Liu, and Xiaofeng Yang. "Deep Learning in Medical Image Registration: A Review." Physics in Medicine & Biology 65, no. 20 (2020): 20TR
- Abbasi, Samaneh, Meysam Tavakoli, Hamid Reza Boveiri, Mohammad Amin Mosleh Shirazi, Raouf Khayami, Hedieh Khorasani, Reza Javidan, and Alireza Mehdizadeh. "Medical Image Registration Using Unsupervised Deep Neural Network: A Scoping Literature Review." Biomedical Signal Processing and Control 73 (2022): 103444.
Point 3: The proposed workflow relies on the segmentation of gray matter and white matter. And Naïve Bayes method was used in the process. Sometimes, Naïve Bayes doesn’t perform very well in several scenarios. Will other segmentation methods, such as U-Net architecture-based methods, help to improve the final performance?
Response 3: We agree with the comment made by the reviewer. This work mainly includes brain tissue segmentation, morphological edge detection of different brain tissues, corner detection, detection of feature points, screening of feature points, and spatial transformation. In this registration process, all steps can be further optimized. As the reviewer described, U-Net architecture-based methods can help to improve the final performance. We have added this relevant contents in the limitations section. This will be one part of our future work plan.
Point 4: The authors claimed the proposed workflow is the closest to the reference image in Figs. 6-9 (lines 290-296). However, it’s visually not easy to tell the difference between the proposed method and SURF, particularly in Figs. 6-8.
Response 4: We thank the reviewer for the comment. Compared with other methods, the results of our method are closer to those of the surf. However, the results of our method are totally different from those of the Surf algorithm in brain contour and spatial location. As shown in the following figures, the most obvious difference is in the brain contour for Figure 6 and in the spatial location (the surf results move up as a whole) for Figure 7. In Figure 8, the surf results are skewed relative to our results.
Point 5: The authors claimed that the proposed method can “maximally” assure the correspondence of different individual brain images for spatial localization and anatomical structures (lines 348-349). It will be helpful to include more evidence to support the “maximal” argument.
Response 5: We apologize for our inappropriate statement. According to the suggestions of the reviewer, we corrected “Therefore, the proposed method can maximally assure the correspondence of different individual brain images for spatial location and anatomical structure.” to “Therefore, the proposed method can guarantee the correspondence of different individual brain images for spatial location and anatomical structure as much as possible.”
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
No further comment.
