3D Vascular Pattern Extraction from Grayscale Volumetric Ultrasound Images for Biometric Recognition Purposes
Round 1
Reviewer 1 Report
In this paper, a method for 3D vascular pattern extraction from grayscale volumetric ultrasound images is proposed for biometric recognition, which can be collected in a short time. This paper is well organized and the experimental results declare that it has positive performances. However, I still have some concerns before it can be accepted as follows:
1) The literature review section should be improved with more state-of-the-art related methods published in these two years.
2) In Section 3, why SRAD is used to reduce speckle noise in the image. Please give more detailed explanation.
3) The introduced mean value and Ridler’s threshold methods are very old in threshold binarization. Are there any more advanced binarization methods published in recent years.
4) The contributions and the novelty of the proposed method cannot be clearly presented in the manuscript, which are suggested to be re-organized.
5) The database were acquired from only 22 volunteers. It is not enough to prove the effectiveness of the proposed method.
Author Response
Comments and Suggestions for Authors
In this paper, a method for 3D vascular pattern extraction from grayscale volumetric ultrasound images is proposed for biometric recognition, which can be collected in a short time. This paper is well organized and the experimental results declare that it has positive performances. However, I still have some concerns before it can be accepted as follows:
We thank the reviewer for his comments and his stimulating suggestions.
1) The literature review section should be improved with more state-of-the-art related methods published in these two years.
In the new version of the paper, we added two more references to recent papers dealing with filtering techniques for ultrasound images, a reference to a recent paper analyzing binarization methods and one more reference to a recent paper presenting palm vein recognition with NIR methods.
2) In Section 3, why SRAD is used to reduce speckle noise in the image. Please give more detailed explanation.
As documented by several references in the paper (in the new version we added two of them) SRAD filter is a well-established filtering technique for reducing speckle noise in ultrasound images. We experienced with this kind of filter in previous works, obtaining good results. As we achieved good results after experimenting with several parameter combinations, we didn’t look for other possibilities. However, this is an ongoing research activity. As we stated in the conclusions we are working to improve the recognition system by trying first of all to extract as many vascular patterns as possible, which will extend the number of analyzed samples in the database. Other filtering methods to reduce speckle noise will be evaluated as well.
3) The introduced mean value and Ridler’s threshold methods are very old in threshold binarization. Are there any more advanced binarization methods published in recent years.
The reviewer is right. We added a reference to a recent paper that analyzes many binarization methods in the new version of the paper. However, similarly to the previous answer, we preferred to use well-established methods.
4) The contributions and the novelty of the proposed method cannot be clearly presented in the manuscript, which are suggested to be re-organized.
In the new version of the paper, we clearly highlight the novelty and important contributions in several points of the manuscript.
5) The database were acquired from only 22 volunteers. It is not enough to prove the effectiveness of the proposed method.
Yes, we are aware of this and we clearly assert that quantitative results should be confirmed by experiments on a wider database. However, some results, i.e., the improvement of recognition rate when SRAD filter and 3D templates are used, seems to be quite consistent.
We better clarify this concept in the new version of the paper
Reviewer 2 Report
Please elaborate on why only a single vascular pattern was extracted and used in the experiments.
The dataset used is a subset of a larger dataset, and it is not clear how the subset is selected. It is best to utilize a bigger sample of the dataset or to use the dataset in its entirety to have results and analyses that are more convincing.
is found in image n+1 then xz coordinates of the one detected in the previous image (n) are assigned. If this condition occurs for j consecutive images, coordinates of missing centroids are calculated through interpolation from those of centroids in images n and n+j. Based on a heuristic analysis of the images in the database, a further crop is performed and only 461 B-mode images were analyzed, from the 200th to 640th."
Overall comment: It seems the proposed approach is highly fine-tuned for the dataset, and lack the generalization criteria for it to be valid for other datasets.Author Response
Comments and Suggestions for Authors
We thank the reviewer for his comments and his stimulating suggestions.
Please elaborate on why only a single vascular pattern was extracted and used in the experiments.
To our best knowledge, this is the first work were hand vein pattern are extracted from greyscale ultrasound images for biometric recognition purposes. So we decided to initially focus on a single pattern to be able to verify extraction accuracy by comparisons with c-mode images extracted at various depths and to gain confidence on the recognition capability by performing preliminary experiments.
We better clarify this concept in the new version of the paper
The dataset used is a subset of a larger dataset, and it is not clear how the subset is selected. It is best to utilize a bigger sample of the dataset or to use the dataset in its entirety to have results and analyses that are more convincing.
As 3D images of the database were collected for extracting palmprint features in a previous work, which are located in the shallower part of the hand palm, many samples are not sufficiently thick (see an example in Figure 11 and the related comments) to contain a full vascular pattern. For the selected vascular pattern only 64 samples were suitable. As we stated in the conclusions, future work will be devoted to extract all vascular patterns contained in the volume, which will increase the number of samples.
We better clarify this concept in the new version of the paper
Please find below my elaborated comments.
- It is mentioned that the best results using the SRAD filter are achieved using a selection of parameters (e.g. 50 x 50 window). However, there is no notion of what metric is used to determine the best results for the selection of these parameters. Moreover, it is not clear whether these parameters have the same effect on all of the images.
In this work, the optimum set of parameters for SRAD filter was determined in a heuristic way, we choose parameters that maximized recognition results. We better clarify this concept in the new version of the paper.
- It seems the pre-processing approach is highly customized for the dataset. "The selection of the range of the connected components that are a candidate to be recognized as the vein is first performed by removing areas too wide (> 1600) and too small (< 129)." How can we scale this approach for other datasets? And will the range selection have an adverse effect on the final results obtained? Why not use a standard connected component analysis approach? Please justify.
The specific values 1600 and 129 were chosen in a heuristic way, by estimating the maximum and minimum dimension of the vein and verifying the assumption by comparing the patterns with C-mode images. Upper and a lower limits have to be fixed to eliminate connected regions that are not related to vein. In fact, as can be seen in Figure 5, there are connected regions too wide (see for example in the top and the bottom of the image) or too small (almost everywhere) to be considered as a vein.
These considerations were made by analyzing only the images of our database because, to our best knowledge, there is no other dataset available. We do not exclude that these values could be slightly modified by considering a different or broader database.
We better clarify this concept in the new version of the paper
- The method is highly customized for the dataset. The following paragraph presents a customized approach based on heuristics that makes the entire approach seems rather hard to be generalized.
"To hold this problem, the following strategy has been devised and experimented with: only centroids that are no more than 12 pixels away on the z-axis and 5 on the x-axis from the reference centroid of the previous image are considered valid.
Actually, these values were chosen based on the specific images analyzed.
If no centroid is found in image n+1 then xz coordinates of the one detected in the previous image (n) are assigned. If this condition occurs for j consecutive images, coordinates of missing centroids are calculated through interpolation from those of centroids in images n and n+j.
This is a general interpolation to assign a value in the case of missing centroids.
Based on a heuristic analysis of the images in the database, a further crop is performed and only 461 B-mode images were analyzed, from the 200th to 640th."
This crop is performed to select the volume where veins are clearly detectable.
Overall comment: It seems the proposed approach is highly fine-tuned for the dataset, and lack the generalization criteria for it to be valid for other datasets.
As said before, there is no other similar database at the moment in the literature. We agree that when other (hopefully larger) databases will be available a more general approach would be desirable.
We better discuss these concepts in the new version of the paper.
Reviewer 3 Report
The paper highlights the benefits of using 3D vascular patterns from greyscale images of the human hand. Some verification and identification tests are performed. The results of the tests are reported to be EER of 2%, ROC AUC of 99.92% and identification rate of 100%.
1. The merit of the results obtained is unclear as the evaluation tests are performed on a local small database that has 64 acquisitions from 22 volunteers. The database from the previous work itself has 633 samples from 136 subjects. It is unclear why the whole database was not considered for the tests. An identification rate of 100% (reported) is possible in a small test case but may fail in a larger and more realistic setting. Further testing needs to be done to verify if the results are consistent.
2. "On the other hand, NIR images may erroneously interpreter some visible skin features as vascular patterns". If the authors mean ' erroneously interpret' then this statement could be challenged. Vascular NIR images capture the vein patterns. Most vein databases provide vein images (i.e. NIR images) and also images under normal light eg. HK PolyU Multimodal Palm database. Any extra information (visible skin features) in the acquisition stage can be used as an added feature as it can be unique to the person and can in turn be used for the recognition process.
3. The authors need to consider the practicality of the acquisition process or defend the benefit of their approach strongly. This has not been done.
"The acquisition procedure involves the user dipping his hand in a basin of water with the palm facing upwards while 57 the probe, also immersed in water, is mechanically moved in the direction from the wrist to the end of 58the palm (fingers excluded), while the system continuously acquires B-mode images"
In a post-pandemic world, how relevant/feasible would the above-mentioned process be? Further, there is no information about the cost analysis for the experimental setup which is a major concern in current vascular biometric systems. Biometric systems are moving towards contactless approaches. If a contact-based method is being suggested, its merits should be clearly highlighted which is not done in this paper.
4. The feature extraction process suggests heavy post-processing to obtain a binarised version of the acquired image. The authors should look into much simpler methods such as using Gabor filters or Curvelet transforms. Further, there are many deep learning-based segmentation models that can extract features accurately. The authors should investigate if these algorithms could be used and comment on the relevant analysis/observations.
5. The verification and identification test setup are premature. The authors should consider a wider experimental setup and then report more specific details with standardized performance metrics such as Precision, Recall, F1-Score that are widely used in this field of research.
6. The paper has many spelling and grammatical errors. This should be worked on.
Author Response
Comments and Suggestions for Authors
The paper highlights the benefits of using 3D vascular patterns from greyscale images of the human hand. Some verification and identification tests are performed. The results of the tests are reported to be EER of 2%, ROC AUC of 99.92% and identification rate of 100%.
We thank the reviewer for his comments and his stimulating suggestions.
- The merit of the results obtained is unclear as the evaluation tests are performed on a local small database that has 64 acquisitions from 22 volunteers. The database from the previous work itself has 633 samples from 136 subjects. It is unclear why the whole database was not considered for the tests. An identification rate of 100% (reported) is possible in a small test case but may fail in a larger and more realistic setting. Further testing needs to be done to verify if the results are consistent.
To our best knowledge, this is the first work were hand vein pattern are extracted from greyscale ultrasound images for biometric recognition purposes and no other database can be found in the literature. So we decided to initially focus on a single pattern to be able to verify extraction accuracy by comparisons with c-mode images extracted at various depths and to gain confidence on the recognition capability by performing preliminary experiments.
As 3D images of the database were collected for extracting palmprint features in a previous work, which are located in the shallower part of the hand palm, many samples are not sufficiently thick (see an example in Figure 11 and the related comments) to contain a full vascular pattern. For the selected vascular pattern only 64 samples were suitable. As we stated in the conclusions future work will be devoted to extract all vascular pattern contained in the volume, which will increase the number of samples.
We better clarify this concept in the new version of the paper
- "On the other hand, NIR images may erroneously interpreter some visible skin features as vascular patterns". If the authors mean ' erroneously interpret' then this statement could be challenged. Vascular NIR images capture the vein patterns. Most vein databases provide vein images (i.e. NIR images) and also images under normal light eg. HK PolyU Multimodal Palm database. Any extra information (visible skin features) in the acquisition stage can be used as an added feature as it can be unique to the person and can in turn be used for the recognition process.
We are not experts on NIR images so we can’t question reviewer observation. In the new version of the paper, we just report the sentence taken from the reference (faces the problem of pattern corruption because of visible skin features being mistaken for veins) in quotes.
- The authors need to consider the practicality of the acquisition process or defend the benefit of their approach strongly. This has not been done.
"The acquisition procedure involves the user dipping his hand in a basin of water with the palm facing upwards while the probe, also immersed in water, is mechanically moved in the direction from the wrist to the end of 58the palm (fingers excluded), while the system continuously acquires B-mode images"
In a post-pandemic world, how relevant/feasible would the above-mentioned process be? Further, there is no information about the cost analysis for the experimental setup which is a major concern in current vascular biometric systems. Biometric systems are moving towards contactless approaches. If a contact-based method is being suggested, its merits should be clearly highlighted which is not done in this paper.
The setup used for collecting 3D images uses water as coupling medium between probe and human hand. Water is often used for preliminary testing of new ultrasound images acquisition systems because it is cheap and performs slightly better than gel, which is the other alternative. As stated in the conclusions, a feasible and practical acquisition system based on gel (which could be acceptable in a post-pandemic world) has already proposed and tested for palmprint extraction (references are given in the paper) and could be exploited for vein extraction as well.
We better explained these concepts in the new version of the paper
- The feature extraction process suggests heavy post-processing to obtain a binarised version of the acquired image. The authors should look into much simpler methods such as using Gabor filters or Curvelet transforms. Further, there are many deep learning-based segmentation models that can extract features accurately. The authors should investigate if these algorithms could be used and comment on the relevant analysis/observations.
We agree with the reviewer that the extraction method may be not optimal, but it works. As said before, future work will be devoted to extract all vascular patterns contained in the volume, which will increase the number of samples. We will for sure investigate other possibilities, including the use of machine learning methods, which however need a wider database for training.
- The verification and identification test setup are premature. The authors should consider a wider experimental setup and then report more specific details with standardized performance metrics such as Precision, Recall, F1-Score that are widely used in this field of research.
As already mentioned, this is the first work were vascular pattern are extracted from greyscale ultrasound images for biometric recognition purposes. Even if results were achieved by considering only one vein pattern and, consequently, a relatively low database, which limits its reliability, they provide a first validation of this method. In particular, it is demonstrated that 3D vascular patterns, which are a peculiarity of Ultrasound over infrared techniques, are able to provide better recognition accuracy than 2D ones. More accurate and detailed analyses will be done in future works, including methodology suggested by the reviewer.
We better explained these concepts in the new version of the paper
- The paper has many spelling and grammatical errors. This should be worked on.
We carefully checked the work.
Round 2
Reviewer 2 Report
Sufficient changes were made. However, as mentioned in my earlier review, the proposed approach is highly tailored to the dataset, using a heuristic approach in several key steps in the proposed approach.
In future work, please consider automating these heuristic steps.
Author Response
Acknowledged.
Reviewer 3 Report
1. To our best knowledge, this is the first work were hand vein pattern are extracted from greyscale ultrasound images for biometric recognition purposes and no other database can be found in the literature. So we decided to initially focus on a single pattern to be able to verify extraction accuracy by comparisons with c-mode images extracted at various depths and to gain confidence on the recognition capability by performing preliminary experiments. As 3D images of the database were collected for extracting palmprint features in a previous work, which are located in the shallower part of the hand palm, many samples are not sufficiently thick (see an example in Figure 11 and the related comments) to contain a full vascular pattern. For the selected vascular pattern only 64 samples were suitable. As we stated in the conclusions future work will be devoted to extract all vascular pattern contained in the volume, which will increase the number of samples.
The revised version of the paper accounts for the change and is much clear with the methodology taken into account.
2. We are not experts on NIR images so we can’t question reviewer observation. In the new version of the paper, we just report the sentence taken from the reference (faces the problem of pattern corruption because of visible skin features being mistaken for veins) in quotes.
This is now in a reasonable form to be of value to the wide research community.
3. The setup used for collecting 3D images uses water as coupling medium between probe and human hand. Water is often used for preliminary testing of new ultrasound images acquisition systems because it is cheap and performs slightly better than gel, which is the other alternative. As stated in the conclusions, a feasible and practical acquisition system based on gel (which could be acceptable in a post-pandemic world) has already proposed and tested for palmprint extraction (references are given in the paper) and could be exploited for vein extraction as well. We better explained these concepts in the new version of the paper
Although the effectiveness of this acquisition method is still questionable. The process and concept is better explained in the updated version. The justification for the choice can be backed up with more references showing the effectiveness of volumetric ultrasound.
4. We agree with the reviewer that the extraction method may be not optimal, but it works. As said before, future work will be devoted to extract all vascular patterns contained in the volume, which will increase the number of samples. We will for sure investigate other possibilities, including the use of machine learning methods, which however need a wider database for training.
Investigating machine learning methods for feature extraction will be beneficial. However, the revised version justifies your choice and is now in an acceptable form when looking at the overall idea flow.
Author Response
We added two references to works on the medical use of 3D US images and one more to the Qualcomm website where the benefits of 3D ultrasound in biometric recognition applications are illustrated.
