Submit to this Journal Review for this Journal Propose a Special Issue

Article Menu

Printed Edition

A printed edition of this Special Issue is available at MDPI Books....

Share Help Cite Discuss in SciProfiles

Open AccessArticle

Peer-Review Record

Automatic Breast Tumor Screening of Mammographic Images with Optimal Convolutional Neural Network

Appl. Sci. 2022, 12(8), 4079; https://doi.org/10.3390/app12084079

by Pi-Yun Chen¹, Xuan-Hao Zhang¹, Jian-Xing Wu¹, Ching-Chou Pai^1,2, Jin-Chyr Hsu³, Chia-Hung Lin^1,*

and Neng-Sheng Pai^1,*

Reviewer 1: Anonymous

Reviewer 2: Anonymous

Reviewer 3:

Ibomoiye Domor Mienye

Reviewer 4:

Alessandro Bruno

Appl. Sci. 2022, 12(8), 4079; https://doi.org/10.3390/app12084079

Submission received: 13 February 2022 / Revised: 3 April 2022 / Accepted: 15 April 2022 / Published: 18 April 2022

(This article belongs to the Special Issue Advances in Biomedical Image Processing and Analysis)

Round 1

Reviewer 1 Report

Dear authors,

After reviewing your manuscript, I found it interesting and I think it can be interesting to the Journal's readers. However, minor to moderate revisions are required. Please see the attached word document.

Kind regards,

Reviewer

Comments for author File: Comments.pdf

Author Response

For Reviewer: #1

Response: The point-to-point responses to all the referees are shown below.

Title: The title clearly describes the subject of the paper.

Response: Thank you for reviewer’s comment. The paper title has been modified as “Automatic Breast Tumor Screening of Mammographic Images with Optimal Convolutional Neural Network” in Page#1.

Abstract and Keywords: Informative. Provides a brief overview on computational techniques. Describes this current study and what is achieved. The results are noted but not clearly specified. This is not an issue. Overall, the abstract and keywords are adequate.

Response: Thank you for reviewer’s comment. Abstract has been slightly modified in Page#1.

Abstract:

… the proposed classifier performs the fractional-order convolutional process to enhance the image and remove unwanted noise for obtaining the desired object’s edges; in the second and third convolutional-pooling layers, two kernel convolutional and pooling operations for the continuous enhancement and sharpening feature patterns for further extracting the desired features at different scales and different levels and also reducing the dimensions of the feature patterns. …

Section 1: The Introduction section: • addresses breast cancer, breast cancer cases, probability, symptoms, and pattern of detection • artificial intelligence and Big Data are noted from aspect of automatic tumor screening • machine learning, deep learning, CNN networks are noted and discussed • CNN-based methods are explained in more detail • the architecture of the proposed multilayer CNN for automatic breast tumor screening is presented I recommend that a brief paragraph on the existing gap in literature is noted, and how this study aims to fill this gap. Furthermore, consider splitting the Introduction section into another separate section titled "Theoretical background" or "Research framework". This would reduce the size of the Introduction section, and readers could assess the theoretical aspects in the following section. In addition, newer research, from 2021 can be introduced in order to expand the literature overview.

Response: Thank you for reminding us. Introduction has been modified and some references have been added in Introduction in Pages#2 and #3.

Section 2: This section provides adequate detail and information on the research methodology. The positive and new aspects of the designed multilayer deep-learning-based CNN could be briefly noted.

Response: Thank you for reviewer’s comment.

Section 3: The experimental setup should be noted in the previous section. This section should contain only the obtained results. Please consider restructuring.

Response: Thank you for reviewer’s comment. The experimental setup has been moved to Section 2.5.

This should be two separate sections. The Discussion section could include additional previous studies and comment on the limitations and advantages of this current study compared to previous studies. If no comparisons are possible, discuss why, and how this current study exceeds previous approaches. The Conclusion section should be a brief, concise overview on the obtained results and conducted study. Suggestions, guidelines and possible interesting future studies could be concisely discussed.

Response: Thank you for reviewer’s comment. The Discussion and Conclusion Sections have been restricted, in Pages#13 and #14.

Discussion:

… 3.4. Discussion

This study designs a mammography classification method incorporating a multilayer CNN-based classifier for automatic breast tumor screening in clinical applications. The proposed classifier algorithm is implemented in the LabVIEW 2019 (NI^TM) software, MATLAB Script tools, and open source Tensorflow platform (Version 1.9.0) [28] and integrated into a computer assistive system with the automatic and manual feature extraction and breast tumor screening modes. The fractional-order convolutional layer and two convolutional-pooling layers allow the image enhancement, and sharpening of the possible tumor edges, contours and shapes via one fractional-order and two kernel convolutional processes in the feature patterns. Through a series of convolution and pooling processes at different scales and different dimensions, the classifier can obtain nonlinearity feature representation from low-level features to high-level information [29]. Then, with the specific bounding boxes (automatic or manual mode) for ROI extraction, enhanced feature patterns can then be distinguished for further breast tumor screening by the multilayer classifier in the classification layer. A gradient-descent optimization method, namely, the ADAM algorithm, is used in the back-propagation process to adjust the network weighted parameters in the classification layer. With K-fold (K_f=10) cross-validation, randomly selected 466 untrained feature patterns for each test fold, the proposed multilayer CNN-based classifier has high recall (%), precision (%), accuracy (%), and F1 scores for screening abnormalitie in both right and left breasts. Experimental results show that the proposed multilayer CNN model offers image enhancement, feature extraction, automatic screening capability, and higher average accuracy (larger than 95%) for separating the normal condition from the possible tumor classes. In literatures [3-7, 10, 56], multilayer CNNs comprised several convolutional - pooling layers and a fully connecting network to establish a classifier for automatic breast tumor screening and could also applied for CT, MRI, chest X-ray, ultrasound image processes, such as image classification and segmentation in clinical applications [19, 23, 28, 35-36, 52, 55]. The combination of cascade of deep liraning and a fully connecting network is also carried out a multilayer CNN based classifier and a decision scheme [56]; for screened suspicious region on mammograms, the cascade of deep liraning method had 98% of sensitivity and 90% of specificity on SuReMapp (Suspicious Region Detection on Mammogram from PP ) dataset [57] and 94% of sensitivity and 91% of specificity on mini-MIAS dataset [56]. This CNN based multilayer classifier could extract multi-scale feature patterns and increase the depth and width feature patterns by using multi convolutional – pooling processes, resuled in accuracy increases. However, over much multi convolutional processes would completely lose internal data about the position and the orientation of the desired object, and over much multi pooling processing would lose va;uable information relative spatial relationships between features, thus, much processes were required to perform with GPU hardware for complexity computational processess. Hence, the proposed optimal multilayer CNN architecture could contain 2D spatial information in fractional-order convolutional layer, and then continuously enhanced the features with two-round convolutional–pooling processes, which could extract the desired features at different scales and different levels. Thus, comparision with the other deep learning methods, the proposed multilayer classifier had also promising results for reached the desired medical purpose. Hence, we have some advantages for proposed CNN-based classifier, as following: …

Conclusion:

… 4. Conclusion

The proposed CNN architecture had better learning ability for complex feature patterns in massive-sized training datasets, and also had promsing classifier’s performance than traditional CNN based classifiers and cascade of deep liraning based classifiers. Through experimental test and validation, we suggested optimal architecture for simplified and established a multilayer CNN based classifier, consisted of a fractional-order convolutional layer, two Kernel convolutional – pooling layers, and a classification layer. Hence, this optimal CNN based classifie could replace manual screening for tasks requiring specific expertise and experience for medical image examination, which could also raise its indication in clinical applications with CBIS-DDSM and SuReMapp dataset for trained the proposed classifier. Additionally, in real-world applications, clinical mammography with biomarker are continuously obtained, the new feature patterns can be extracted and added to the current database to further train the CNN-based classifier, which can keep its intended medical purpose and can also be used as a computer-aided decision-making tool and a software in a medical device tool. …

Adequate. Includes relevant literature sources. However, as the discussion section is suggested to be expanded, newer literature sources should be addressed (from the year 2021).

Response: Thank you for reminding us. The manuscript has been modified with blue highlighting indicating changes, and some references (years 2018 - 2021) have been added in the manuscript.

Author Response File: Author Response.doc

Reviewer 2 Report

Journal: Applied Sciences

Title: Mammography Classification with Optimal Convolutional Neural Network for Automatic Breast Tumor Screening

Pi-Yun Chen, Xuan-Hao Zhang, Jian-Xing Wu, Ching Chou Pai, Jin-Chyr Hsu, Chia-Hung Lin*, Neng-Sheng Pai

The authors attempted to develop an automatic breast cancer recognition algorithm. Four performance parameters have been applied and tenfold cross-validation were introduced to evaluation of classifier’s performance. The dataset was randomly divided in two equal halves: 50% training- and 50% test sets. Noise removal has been included by a “fractional-order convolutional process” whatever it is.

To reduce the 2D image of 4,320 pixels × 2,600 pixels to 100 pixels × 100 pixels and even further to 9 × 9 matrices realizes a very large information loss. Alone this fact seriously questions the plausibility of the findings.

Another fault is the lack of reproducibility. I doubt that anybody can reproduce the investigations solely on the basis of the description (including the authors themselves).

The alignment problem has not even mentioned, although it is well-known that a small difference in the alignment can seriously distort the performance.

The validation is also problematic. The tenfold cross-validation is important, a necessary but not sufficient condition. Randomization test (y-scrambling), external validation and above all, independent new samples should be delivered to the classification algorithms. The results should be compared with experts” classifications.

Table 1 is, in fact, a contingency table for binary classification. Figure 2 involves three classes. Anyway, the number of performance parameters are too low to be drawn any reasonable conclusions. A summary of binary performance merits are gathered in Molecules 2019, 24, 2811; doi:10.3390/molecules24152811 .

Minor errors

The title is imprecise: “Mammography Classification” would mean a classification of apparatuses and not tumors.
“156 mammography images (78 subjects)” – repetitions or/and both breast?
“2,500 enrolled subjects … containing normal cases, benign cases, malignant cases, … which provides ground truth validation” – Come on! “Ground truth” does not exist. Nothing is said about the balancing of the classes. Otherwise, a simple biased classification would produce high performance parameters.
“This reduction in the number of parameters that must be trained overcomes the overfitting problem” – by no means! Feature reduction can diminish the probability of wrong decision, but overfitting is inherent in the convolutional neural network technique.
It is not clear whether the cross-validation has been performed repeatedly (with return) or not.
Eqs (2) and (11) consists of two equations.
- The discussion and conclusion are rather a summary and not a discussion of classical art and not concluding remarks. The conclusion should summarize what follows form the investigations, what is the consequence of the classification model(s).
A discussion should discriminate the present and earlier findings and compare the performance with earlier approaches.
Abbreviations are to be avoided in titles, abstract and conclusions.
, etc.

I have enumerated some problems above without being complete. There is no wonder that the manuscript cannot be accepted in its present from. Personally, I do not believe that the authors can improve it to be acceptable. Reproducible description, correct validation, convincing performance parameters, valid discussion emphasizing the novelty, correction of minor errors are all necessary to get an accepted manuscript.

February 21 / 2021 referee:

Author Response

For Reviewer: #2

Response: The point-to-point responses to all the referees are shown below.

To reduce the 2D image of 4,320 pixels × 2,600 pixels to 100 pixels × 100 pixels and even further to 9 × 9 matrices realizes a very large information loss. Alone this fact seriously questions the plausibility of the findings.

Response: Thank you for reminding us. Some sentences have been added in Introduction in Page#2.

Introduction:

… Despite its many advantages, however, a deep-learning based CNN presents several drawbacks and limitations, such as the number of convolutional-pooling layers determination, the sizes of convolutional masks assignment (3 ´ 3, 5 ´ 5, 7 ´ 7, 9 ´ 9, 11 ´ 11), the high computational complexity and large-scale dataset requirement for training CNN based classifier, and the poor suitability for real-time applications. In addition, the multi convolutional–pooling processes with different sizes of convolutional masks will realize a very large information loss for feature extraction and will result in increasing complexity level. The multilayer CNN must be performed with a graphics processing unit (GPU) to speed up the training and classification tasks with a large amount of training and testing dataset. …

Another fault is the lack of reproducibility. I doubt that anybody can reproduce the investigations solely on the basis of the description (including the authors themselves). The alignment problem has not even mentioned, although it is well-known that a small difference in the alignment can seriously distort the performance. The validation is also problematic. The tenfold cross-validation is important, a necessary but not sufficient condition. Randomization test (y-scrambling), external validation and above all, independent new samples should be delivered to the classification algorithms. The results should be compared with experts” classifications.

Response: Thank you for reminding us. Some sentences have been added in Introduction in Page#3.

Introduction:

… A total of 78 subjects are selected from the MIAS (Mammographic Image Analysis Society) Digital Mammogram Database (United Kingdom National Breast Screening Program) for experimental analysis. The clinical information is confirmed and agreed upon by expert radiologists for biomarker, such as image size, image category, background tissue, class of abnormality, and severity of abnormality [30-31]; the image database includes a total of 156 mammography images, including 94 normality cases and 62 abnormalities involving benign and malignant cases, the ROIs are extracted by 100×100 bounding box and then 932 feature patterns are extracted by using the proposed convolutional-pooling processes including 564 abnormalities and 368 tumor-free patterns. Using cross-validation, the dataset can be randomly divided into two equal halve: 50% of dataset for training the classifier and 50% of those for evaluating the classifier’s performance. Thus, tenfold cross-validation is used to verify the performances of the proposed multilayer deep learning-based CNN with the proposed convolutional-pooling layers in terms of recall (%), precision (%), accuracy (%), F1 score, and Youden’s index. Therefore, the optimal architecture of multilayer CNN can be determine and may potentially be applied to establish a classifier for automatic breast tumor screening in clinical applications. …

Table 1 is, in fact, a contingency table for binary classification. Figure 2 involves three classes. Anyway, the number of performance parameters are too low to be drawn any reasonable conclusions. A summary of binary performance merits are gathered in Molecules 2019, 24, 2811; .

Response: Thank you for reminding us. Figure 2 has been modified in Page#4 and one reference has been added in Introduction (doi:10.3390/molecules24152811).

# Minor Errors:

The title is imprecise: “Mammography Classification” would mean a classification of apparatuses and not tumors.

Response: Thank you for reminding us. The paper title has been modified as “Automatic Breast Tumor Screening of Mammographic Images with Optimal Convolutional Neural Network” in Page#1.

“156 mammography images (78 subjects)” – repetitions or/and both breast?

Response: Thank you for reminding us. Some sentences have been modified in Introduction, in Page#3.

Introduction:

“2,500 enrolled subjects … containing normal cases, benign cases, malignant cases, … which provides ground truth validation” – Come on! “Ground truth” does not exist. Nothing is said about the balancing of the classes. Otherwise, a simple biased classification would produce high performance parameters.

Response: Thank you for reminding us. Some sentences have been modified in Introduction, in Pages#2 and #3.

Introduction:

… Additionally, such as CBIS-DDSM (Curated Breast Imaging Subset of Digital Database for Screening Mammography), it is a database of approximately 2,500 enrolled subjects for studies in mammographic classification of breast lesions, including normal cases, benign cases, malignant cases, and pathology information [10], which provides ground truth validation to make the DDSM for applying to develop and validate the decision support systems. …

“This reduction in the number of parameters that must be trained overcomes the overfitting problem” – by no means! Feature reduction can diminish the probability of wrong decision, but overfitting is inherent in the convolutional neural network technique.

Response: Thank you for reminding us. Some sentences have been modified in Section 2.1., in Page#5.

Section 2.1.:

… After MP, the number of feature patterns is reduced to 25% of the total number of original feature images. This reduction in the dimensions of the feature patterns that can overcome the overfitting problem for

training a multilayer classifier. …

It is not clear whether the cross-validation has been performed repeatedly (with return) or not.

Response: Thank you for reminding us. Some sentences have been added in Section 2.5., in Page#7.

Section 2.5.:

… In total, 932 feature patterns, including 564 tumor and 368 tumor-free screenshots, are obtained. In each classifier’s training stage, 282 tumor and 184 tumor-free screenshots (50% feature patterns) are randomly selected to train the multilayer CNN classifier. The remaining 50% of the feature patterns are used to evaluate the classifier’s performance for each cross validation. …

Eqs (2) and (11) consists of two equations.

Response: Thank you for reminding us. Yes, Eqs (2) and (11) consist of two equations.

The discussion and conclusion are rather a summary and not a discussion of classical art and not concluding remarks. The conclusion should summarize what follows form the investigations, what is the consequence of the classification model(s).

Response: Thank you for reviewer’s comment. The Discussion and Conclusion Sections have been restricted, in Pages#13 and #14.

Discussion:

… 3.4. Discussion

Conclusion:

… 4. Conclusion

A discussion should discriminate the present and earlier findings and compare the performance with earlier approaches.

Response: Thank you for reviewer’s comment. The Discussion Section has been restricted, in Page#13.

Discussion:

… 3.4. Discussion

Abbreviations are to be avoided in titles, abstract and conclusions., etc.

Response: Thank you for reminding us.

I have enumerated some problems above without being complete. There is no wonder that the manuscript cannot be accepted in its present from. Personally, I do not believe that the authors can improve it to be acceptable. Reproducible description, correct validation, convincing performance parameters, valid discussion emphasizing the novelty, correction of minor errors are all necessary to get an accepted manuscript.

Response: Thank you for reminding us. The manuscript has been modified with blue highlighting indicating changes

Author Response File: Author Response.doc

Reviewer 3 Report

The study proposed a multilayer deep learning architecture to detect breast cancer and solve some of the limitations in the traditional CNN architecture. The article is well written and innovative. I only have a few minor comments:

In lines 27 - 29, it is not clear what the second and third convolutional-pooling layers do, kindly complete the sentence or paraphrase it to make it clearer. "in the second and third convolutional-pooling layers, two kernel convolutional and pooling operations for the continuous enhancement, sharpening, and extraction of feature patterns"
Though the assertions in the paragraph of Lines 124-134 seems correct, I think they should be backed up by suitable references.
Table 8 seems confusing, is the Table showing the results of 'k-fold cross validation"?, if yes, then replace the "ten-fold" in the Table's title with "k-fold". Meanwhile, the F1-score of the 6-fold does not seem like the harmonic mean of the Precision and Recall values, kindly check and correct it.
Lastly, kindly proofread the manuscript for some minor grammatical errors, and formatting issues such as the equation numbers, which should be on the far right.

Author Response

For Reviewer: #3

Response: The point-to-point responses to all the referees are shown below.

In lines 27 - 29, it is not clear what the second and third convolutional-pooling layers do, kindly complete the sentence or paraphrase it to make it clearer. “in the second and third convolutional-pooling layers, two kernel convolutional and pooling operations for the continuous enhancement, sharpening, and extraction of feature patterns”

Response: Thank you for reminding us. Some sentences have been added in Abstract, in Page#1.

Abstract:

Though the assertions in the paragraph of Lines 124-134 seems correct, I think they should be backed up by suitable references.

Response: Thank you for reminding us. The paragraph has been modified in Introduction, in Page#2.

Introduction:

… The 2D CNNs may consist of several convolutional-pooling layers and a fully connected network in the classification layer, such as support vector machines, back-propagation neural networks, and Bayesian networks, which combine the image enhancement, feature extraction and classification tasks into an individual scheme [16-17] to achieve promising accuracy for image classification in breast tumor screening. These CNNs are usually greater than 10 convolutional - pooling layers to perform the abovementioned image preprocessing and postprocessing tasks and then increase the identification accuracy. Hence, this multilayer design may gradually replace machine learning (ML) methods [18-19], which perform the image segmentation and feature extraction as an image preprocess for mammograms and breast MRIs and then use the fixed features obtained to train a classifier. Both CNN and ML based image segmentation [20] can learn the specific features or knowledge representations to automatically identify the boundaries of ROI and then to detect the breast lesions. Traditional ML methods have fewer parameters that can easily be optimized by the gradient descent optimization or back-propagation algorithms through training with small-to-medium-sized datasets [21-22]. Through a series of convolutional and pooling processes, the multilayer CNN can enhance and extract the desired object at different scales and different levels from low-level features (extract object’s edge) to high-level information (extract object’s shape) for detecting nonlinear features, which can increase nonlinearity and obtain feature representation. Then, the pooling process with maximum-pooling (MP) is used to reduce the sizes of feature maps for obtaining abstract features. Thus, in contrast to the traditional machine-learning method, CNN-based methods can learn to extract the feature patterns from the raw data and improve the classification accuracy significantly. However, small- or medium-sized datasets are insufficient to train a deep-learning based CNN. For example, in literatures [23-26], such as AlexNet (8-layer CNN) [25] and ZFnet [26], the deep learning CNN requires several convolutional-pooling layers and fully connected layers for the large-scale image classification (ImageNet image database [27-28]). This CNN can learn to optimize features during the training stage, process large inputs with sparsely connected weights, adapt to different sizes of 2D images, and reduce error rates. In addition, this approach demonstrates greater computational efficiency compared with the traditional fully connected multilayer perceptron (MLP) networks. Despite its many advantages, however, a deep-learning based CNN presents several drawbacks and limitations, such as the number of convolutional-pooling layers determination, the sizes of convolutional masks assignment (3 ´ 3, 5 ´ 5, 7 ´ 7, 9 ´ 9, 11 ´ 11), the high computational complexity and large-scale dataset requirement for training CNN based classifier, and the poor suitability for real-time applications. In addition, the multi convolutional–pooling processes with different sizes of convolutional masks will realize a very large information loss for feature extraction and will result in increasing complexity level. The multilayer CNN must be performed with a graphics processing unit (GPU) to speed up the training and classification tasks with a large amount of training and testing dataset. …

Table 8 seems confusing, is the Table showing the results of 'k-fold cross validation"?, if yes, then replace the "ten-fold" in the Table's title with "k-fold". Meanwhile, the F1-score of the 6-fold does not seem like the harmonic mean of the Precision and Recall values, kindly check and correct it.

Response: Thank you for reminding us. Table 8 has been modified.

Lastly, kindly proofread the manuscript for some minor grammatical errors, and formatting issues such as the equation numbers, which should be on the far right.

Response: Thank you for reminding us. Correct as suggestion.

Author Response File: Author Response.doc

Reviewer 4 Report

Overall, I think your contribution shows some somewhat interesting results regarding the effectiveness of the task supported by high accuracy rates.

Here are some comments below:

You should refurbish the first section by including only elements to introduce the main topic and the main contributions of your work.

If you prefer not to have a related technique section, I guess you should add some references in section 1.

Speaking of breast cancers, some image processing-based techniques are to be added to widen the scope of your introductory section.

For instance, some methods work out preprocessing steps such as breast region segmentation using region growing techniques as in the following two papers:

https://www.sciencedirect.com/science/article/pii/S1877750317310062?casa_token=sBhmWcv0vGMAAAAA:xDxwi1npH7-PlMK3MT18iShVwf94524hf2_kRP0gG4GI0DmJodkD_pCnmE_S5B7xWqdBPUvT

https://link.springer.com/chapter/10.1007/978-3-319-68560-1_63

Regarding the structure of your paper. I think it needs some adjustments. Provide further details of the main components of your proposed method (as you currently do in section 1). For instance, section 2 should be structured as follows:

Main components of the proposed method;

Dataset (description of the dataset employed for the experiments)

Training (number of epochs, stride parameter, all details regarding backpropagation and activation function)

You also describe some feature extraction steps that reportedly are accomplished to improve the overall accuracy of the system. Yet, it is not that clear whether the deep learning solution you propose embeds them or if they are carried out as pre-processing steps.

Regarding the experiments and conclusion section, I think your paper would benefit from having comparison methods. For instance, in the following work, there is a table listing the performances of both traditional and deep learning techniques on MIAS and a SuReMaPP dataset. I think it should be part of your reference articles.

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7528986/

You can easily add a table to your paper to show how your method ranks.

Some minor changes follow down below:

Line 44: Since you already started the previous paragraph with “as per statistics”, I would suggest replacing it with “As per latest figures” or “According to the latest figures”.

Line 46: There must be a typo in the line. Did you mean to state “they standardized”?

From line 46 onward: To have your paper compliant with the MDPI standards, I invite you to replace operational operators such as > < with “larger than” and “lower than”.

Line 54: I appreciate the authors providing the first section with a thorough description of symptoms. Nevertheless, I would recommend adding at least a reference to corroborate what is stated in lines 51-56.

Line 58: I would rephrase this statement using broader terms. For instance, when radiologists investigate biomedical images, they might want to investigate further some regions of interest to get more features and avoid false negatives.

Lines 61-65: please add a scientific reference to ground the stats and figures depicted.

Lines 74-80: Could you please rephrase the statement? You might want to break it down into two shorter sentences.

Line 79: The statement starting at the end of line 79 pertains to the previous sentence. Please fix this accordingly.

Line 101: Please revise the statement in line 101. Are you sure that CNNs may include support vector machines? I think it is not theoretically correct.

Line 118. Please rephrase it; there are some syntax errors.

Line 143. max-pooling rather the maxi-mum pooling

From line 160 onwards in section 1, I suggest you refrain from adding too many technical details in the introductory section. For instance, authors may want to move some technical elements regarding the proposed changes to CNN to the proposed method section. Therefore, the introduction section should provide the primary and overall description of the main contributions.

Please rephrase the statement in line 201.

Author Response

For Reviewer: #4

I appreciate you tackled a hot topic such as the detection of suspicious regions in mammograms. You proposed a method that mostly relies on Convolutional Neural Networks with some changes to some components in the architecture. Overall, I think your contribution shows some somewhat interesting results regarding the effectiveness of the task supported by high accuracy rates.

Response: The point-to-point responses to all the referees are shown below.

#Here are some comments below:

You should refurbish the first section by including only elements to introduce the main topic and the main contributions of your work. If you prefer not to have a related technique section, I guess you should add some references in section 1. Speaking of breast cancers, some image processing-based techniques are to be added to widen the scope of your introductory section. For instance, some methods work out preprocessing steps such as breast region segmentation using region growing techniques as in the following two papers:

https://www.sciencedirect.com/science/article/pii/S1877750317310062?casa_token=sBhmWcv0vGMAAAAA:xDxwi1npH7PlMK3MT18iShVwf94524hf2_kRP0gG4GI0DmJodkD_pCnmE_S5B7xWqdBUvT

https://link.springer.com/chapter/10.1007/978-3-319-68560-1_63

Response: Thank you for reminding us. Introduction has been modified and some references have been added in Introduction in Pages#2 and #3.

Regarding the structure of your paper. I think it needs some adjustments. Provide further details of the main components of your proposed method (as you currently do in section 1). For instance, section 2 should be structured as follows: Main components of the proposed method; Dataset (description of the dataset employed for the experiments) Training (number of epochs, stride parameter, all details regarding backpropagation and activation function)

Response: Thank you for reminding us. Correct as suggestion. Section 2 has been modified.

You also describe some feature extraction steps that reportedly are accomplished to improve the overall accuracy of the system. Yet, it is not that clear whether the deep learning solution you propose embeds them or if they are carried out as pre-processing steps.

Response: Thank you for reminding us. Some sentences have been added in Introduction and Section 2.1., in Pages#2 and #3 and Pages#4 and #5.

Introduction:

… The 2D CNNs may consist of several convolutional-pooling layers and a fully connected network in the classification layer, such as support vector machines, back-propagation neural networks, and Bayesian networks, which combine the image enhancement, feature extraction and classification tasks into an individual scheme [16-17] to achieve promising accuracy for image classification in breast tumor screening. These CNNs are usually greater than 10 convolutional - pooling layers to perform the abovementioned image preprocessing and postprocessing tasks and then increase the identification accuracy. Hence, this multilayer design may gradually replace machine learning (ML) methods [18-19], which perform the image segmentation and feature extraction as an image preprocess for mammograms and breast MRIs and then use the fixed features obtained to train a classifier. Both CNN and ML based image segmentation [20] can learn the specific features or knowledge representations to automatically identify the boundaries of ROI and then to detect the breast lesions. …

… Hence, to simplify the image processing and classification tasks, this study intends to design a suitable number of convolutional-pooling layers and a classification layer to increase the identification accuracy of image classification for automatic breast tumor screening. As seen in Figure 1, we utilized a multilayer classifier, consisting of a fractional-order convolutional layer, two convolutional-pooling layers, a flattening layer, and a multilayer classifier in the classification layer. In the first convolutional layer, 2D spatial convolutional processe with two 3×3 fractional-order convolutional masks are used to perform the enhancement task and remove unwanted noise from original mammographic image for distinguishing the object’s edges and shapes. In the second and third convolutional layers, sixteen 3×3 kernel convolutional windows are used to subsequently enhance and sharpen the feature patterns twice; hence, the tumor contour could easily be highlighted and distinguished for feature pattern extraction. Then, two MP processes are used to reduce the dimensions of the feature patterns, which makes network training avoid falling in overfitting problem [29-30]. …

Section 2.1.:

… Each fractional-order convolutional mask multiples each element, M(i, j) or M(j, i), by the corresponding input pixel values, I(x, y), and then obtains an enhanced feature pattern containing spatial features in the x-axis and y- axis directions. This 2D spatial convolutional processes act as two low-pass frequency filters [39] and then remove the high-spatial-frequency components from a breast mammograms. …

Regarding the experiments and conclusion section, I think your paper would benefit from having comparison methods. For instance, in the following work, there is a table listing the performances of both traditional and deep learning techniques on MIAS and a SuReMaPP dataset. I think it should be part of your reference articles. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7528986/

Response: Thank you for reviewer’s comment. The Discussion Section has been restricted and one reference

(https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7528986/) has been added in Discussion Section, in Page#13.

Discussion:

… 3.4. Discussion

You can easily add a table to your paper to show how your method ranks.

Response: Thank you for reminding us. Table 1 has been modified which is used to show our method ranks (in Page#5).

#Some minor changes follow down below:

Line 44: Since you already started the previous paragraph with “as per statistics”, I would suggest replacing it with “As per latest figures” or “According to the latest figures”.

Response: Thank you for reminding us. Correct as suggestion.

Line 46: There must be a typo in the line. Did you mean to state “they standardized”?

Response: Thank you for reminding us. Correct as suggestion.

From line 46 onward: To have your paper compliant with the MDPI standards, I invite you to replace operational operators such as > < with “larger than” and “lower than”.

Response: Thank you for reminding us. Correct as suggestion.

Line 54: I appreciate the authors providing the first section with a thorough description of symptoms. Nevertheless, I would recommend adding at least a reference to corroborate what is stated in lines 51-56.

Response: Thank you for reminding us. Reference [1] has been added in Introduction, Page#2.

Line 58: I would rephrase this statement using broader terms. For instance, when radiologists investigate biomedical images, they might want to investigate further some regions of interest to get more features and avoid false negatives.

Response: Thank you for reminding us. The sentence has been modified as “However, it is not immediately clear whether the breast lesion is identified or whether metastasis is happening.”

Lines 61-65: please add a scientific reference to ground the stats and figures depicted.

Response: Thank you for reminding us. Reference [1] has been added in Introduction, Page#2.

Lines 74-80: Could you please rephrase the statement? You might want to break it down into two shorter sentences.

Response: Thank you for reminding us. The paragraph has been modified.

Line 79: The statement starting at the end of line 79 pertains to the previous sentence. Please fix this accordingly.

Response: Thank you for reminding us. The sentence has been modified.

Line 101: Please revise the statement in line 101. Are you sure that CNNs may include support vector machines? I think it is not theoretically correct.

Response: Thank you for reminding us. Correct as suggestion.

Line 118. please rephrase it; there are some syntax errors.

Response: Thank you for reminding us. Correct as suggestion.

Line 143. max-pooling rather the maxi-mum pooling

Response: Thank you for reminding us. Correct as suggestion.

From line 160 onwards in section 1, I suggest you refrain from adding too many technical details in the introductory section. For instance, authors may want to move some technical elements regarding the proposed changes to CNN to the proposed method section. Therefore, the introduction section should provide the primary and overall description of the main contributions.

Response: Thank you for reminding us. The sentence has been modified in Introduction, in Page#3.

Introduction:

… Hence, to simplify the image processing and classification tasks, this study intends to design a suitable number of convolutional-pooling layers and a classification layer to increase the identification accuracy of image classification for automatic breast tumor screening. As seen in Figure 1, we utilized a multilayer classifier, consisting of a fractional-order convolutional layer, two convolutional-pooling layers, a flattening layer, and a multilayer classifier in the classification layer. In the first convolutional layer, 2D spatial convolutional processe with two 3×3 fractional-order convolutional masks are used to perform the enhancement task and remove unwanted noise from original mammographic image for distinguishing the object’s edges and shapes. In the second and third convolutional layers, sixteen 3×3 kernel convolutional windows are used to subsequently enhance and sharpen the feature patterns twice; hence, the tumor contour could easily be highlighted and distinguished for feature pattern extraction. Then, two MP processes are used to reduce the dimensions of the feature patterns, which makes network training avoid falling in overfitting problem [29-30]. In the classification layer, a multilayer classifier with an input layer, two hidden layers, and an output layer is implemented to perform the pattern recognition task, which separates tumor-free feature patterns from tumor feature patterns. To reduce the error rates, an adaptive moment estimation method (ADAM) can compute the adaptive learning rates for updating network parameters by storing an exponentially decaying average of past squared gradients [31-32], which combines two stochastic gradient descent approaches, including adaptive gradients and root mean square propagation. Its optimization algorithm uses randomly selected training data subsets to compute the gradient, instead of using the entire dataset. The momentum term can speed up the gradient descent by converging faster. The ADAM algorithm has simple implementation, computation efficiency, and less memory requirements, and is appropriate for operations with large data sets and parameters for training the multilayer CNN models. A total of 78 subjects are selected from the MIAS (Mammographic Image Analysis Society) Digital Mammogram Database (United Kingdom National Breast Screening Program) for experimental analysis. The clinical information is confirmed and agreed upon by expert radiologists for biomarker, such as image size, image category, background tissue, class of abnormality, and severity of abnormality [33-34]; the image database includes a total of 156 mammography images (including right and left images), including 94 normality cases and 62 abnormalities involving benign and malignant cases, the ROIs are extracted by 100×100 bounding box and then 932 feature patterns are extracted by using the proposed convolutional-pooling processes including 564 abnormalities and 368 tumor-free patterns. Using cross-validation, the dataset can be randomly divided into two equal halve: 50% of dataset for training the classifier and 50% of those for evaluating the classifier’s performance. Thus, tenfold cross-validation is used to verify the performances of the proposed multilayer deep learning-based CNN with the proposed convolutional-pooling layers in terms of recall (%), precision (%), accuracy (%), F1 score, and Youden’s index [35-36]. Therefore, the optimal architecture of multilayer CNN can be determine and may potentially be applied to establish a classifier for automatic breast tumor screening in clinical applications. …

Please rephrase the statement in line 201.

Response: Thank you for reminding us. Correct as suggestion.

Author Response File: Author Response.doc

Round 2

Reviewer 2 Report

However, the validation part remained untouched no more performance parameters were included in the study; randomization test has not been carried out either.

The results part has not been changed either. Ignoring the referees' suggestions is not a good policy. Alone this fact should deserve rejection. However, I regret the authors, and give them one more chance. The MS is not acceptable without proper validations and extended results part.

Author Response

Re: Response to reviewers

For Reviewer: #2

The authors have made substantial changes in the manuscript, especially in the Introduction and in the experimental setup part. Therefore, the manuscript is more reproducible and became (much) better. Minor errors are also corrected. However, the validation part remained untouched no more performance parameters were included in the study; randomization test has not been carried out either. The results part has not been changed either. Ignoring the referees' suggestions is not a good policy. Alone this fact should deserve rejection. However, I regret the authors, and give them one more chance. The MS is not acceptable without proper validations and extended results part.

Response: Thank you for reminding us. Some experimental results and some sentences have been added in Sections 3.3 and 3.4, in Pages#11, #13, and #14.

Section 3.3, Pages#11 and #13,

… Considering the experimental results listed in Table 6, the architecture of Model #1 is selected to establish the screening classifier. After training is completed, 466 untrained feature patterns, including 184 abnormal and 282 normal patterns, are randomly selected from the dataset to validate the performance of the classifier. The experimental results of the classifier produce a visual confusion matrix. The testing result of the abnormal pattern yields TP =178 and FP = 6, while that of the normal pattern yields TN = 269 and FN = 13; these values can be used as variables in Equations (14)–(17) to compute the four evaluation indices of the classifier. In this study, precision (%)=96.74%, recall (%)=93.19%, F1 score = 0.9493, and accuracy (%) =95.92%. Precision (%) is the standard for predicting TP, and recall (%) is the true accuracy of TP. Both indicators may be greater than 80%. Recall (%) is also called the positive predictive value (PPV), which is the so-called TP in the detection case. The general PPV index is larger than 80%, which means the proposed classifier has promising predictive performance. The F1 score fuses the indicators of precision (%) and recall (%), and F1 score larger than 0.9000 generally indicates a good classification model. Youden’s index (YI) is a fusion evaluation index of sensitivity (Sens) and specificity (Spec) [54], which reflects the performance of the classifier for detecting abnormalities and abnormalities. The larger the YI, the better the performance of the classifier for detection and validation and the greater its authenticity. The testing results show YI=91.01% (Sens= 93.19%, Spec=97.82%). Given that all evaluation indicators considered in this work exceed larger than 90%, Model #1 indeed has an architecture that supports good classification accuracy and performance, as seen in the tenfold cross-validation (K_f=10) for averages of precision (%), recall (%), accuracy (%), and F1 score in Table 8. Hence, we suggest Model #1 to carry out a multilayer CNN-based classifier for automatic breast tumor screening. In addition, as seen in Table 9, we also set 4, 8, 16, and 32 Kernel convolutional windows and 4, 8, 16, and 32 maximum pooling windows in second and third convolutional - pooling layers, respectively for establishing four models (Model #1-1 to Model #1-4). With the tenfold cross-validation, for the same trained feature patterns, the average training CPU time of Models #1-1 and #1-2 is less than Model #1-3 with 16 Kernel convolutional windows and 16 maximum pooling windows. It can be seen the Mode #1-4 comprises 32 Kernel convolutional windows and 32 maximum pooling windows will increase the average training CPU time and complex computational processes at each cross-validation. With the tenfold cross-validation, for the same untrained feature patterns, as seen in Table 10, the proposed architecture of multilayer classifier (Model #1-3) has promising classification accuracy and performance in terms of average precision (%), average recall (%), average accuracy (%), and average F1 score. Additionally, the proposed CNN architecture with different convolutional windows in the first convolutional layer, including fractional-order, Sobel (first-order), and Histeq convolutional windows, is used to test the performance of breast tumor screening model. Through the tenfold cross-validation, the CNN classifier with a fractional-order convolutional window in the first convolutional layer, as Model #1 in Table 11, has the better classification accuracy (larger than 95%) than Mode1#2 (larger than85%) and Model#3 (larger than 90%). …

Section 3.4, Page#14,

… Experimental results show that the proposed multilayer CNN model offers image enhancement, feature extraction, automatic screening capability, and higher average accuracy (larger than 95%) for separating the normal condition from the possible tumor classes. It has been observed from previous literature [3-7, 10, 56], that multilayer CNNs comprised several convolutional-pooling layers and a fully connecting network to establish a classifier for automatic breast tumor screening, and could also be applied for CT, MRI, chest X-ray, ultrasound image processes, such as image classification and segmentation in clinical applications [19, 23, 28, 35-36, 52, 55]. The combination of a cascade of deep learning and a fully connecting network is also carried out by a multilayer CNN-based classifier, and a decision scheme [56]. For the screened suspicious region on mammograms, the cascade of deep learning method had 98% of sensitivity and 90% of specificity on SuReMapp (Suspicious Region Detection on Mammogram from PP) dataset [57], 94% of sensitivity, and 91% of specificity on mini-MIAS dataset [56]. This CNN-based multilayer classifier could extract multi-scale feature patterns, and increase the depth and width feature patterns by using multi convolutional–pooling processes, which had an overall increase in accuracy. However, excessive multi convolutional-processes would completely lead to a loss of the internal data about the position and the orientation of the desired object, and an excessive multi pooling processing would lose valuable information relating to spatial relationships between features, thus, many processes were required to perform with GPU hardware for complex computational processes. Hence, the proposed optimal multilayer CNN architecture contained 2D spatial information in the fractional-order convolutional layer (with 2 fractional-order convolutional windows), and continuously enhanced the features with two-round convolutional–pooling processes (with 16 Kernel convolutional windows and 16 maximum pooling windows), which could extract the desired features at different scales and different levels. Thus, in comparison with the other deep learning methods, the proposed multilayer classifier exhibited promising results for the desired medical diagnostic purpose. Hence, we have some advantages for the proposed CNN-based classifier, as follows: …

If you have further questions, please feel free to contact me. Your assistance is very much appreciated.

Best Regards,

Dr. Chia-Hung Lin

E-mail eechl53@ gmail.com

Chia-Hung Lin,

Department of Electrical Engineering,

National Chin-Yi University of Technology,

Taichung City, 41170, Taiwan.

E-mail eechl53@ gmail.com

Author Response File: Author Response.pdf

Round 3

Reviewer 2 Report

The authors have not responded to my questions, have not completed a validation. Earlier I wrote: " ... the validation part remained untouched no more performance parameters were included in the study; randomization test has not been carried out either. The results part has not been changed either. Ignoring the referees' suggestions is not a good policy. Alone this fact should deserve rejection. However, I regret the authors, and give them one more chance. The MS is not acceptable without proper validations and extended results part."

As the authors have not responded and corrected the MS accordingly, my opinion has not been changed: rejection without option to resubmission.

BTW The response (and paper) full of minor errors, and no more iteration is warranted, as more and more errors will be appeared.

Some examples:

abnormalitie

deep liraning

resuled in accuracy

lose va;uable

computational processess

comparision with

clinical mammography with biomarker are continuously

promsing classifier’s performance

CNN can be determine

and many more. A scientific paper should be prepared with much more care and precision.

Obviously, the authors lack the necessary scientific school. They have definitely misunderstood their role in the reviewing process. The referees work for free and try to help you to produce a better, more readable manuscript. The reviewers do not need extended explanations; commonly, they know the issue better than the authors themselves. Alternatively, the manuscript should be adjusted to achieve the highest scientific standards. The response should be brief and concise and to the point.

Author Response

For Reviewer: #2

The authors have not responded to my questions, have not completed a validation. Earlier I wrote: " ... the validation part remained untouched no more performance parameters were included in the study; randomization test has not been carried out either. The results part has not been changed either. Ignoring the referees' suggestions is not a good policy. Alone this fact should deserve rejection. However, I regret the authors, and give them one more chance. The MS is not acceptable without proper validations and extended results part."

As the authors have not responded and corrected the MS accordingly, my opinion has not been changed: rejection without option to resubmission. BTW The response (and paper) full of minor errors, and no more iteration is warranted, as more and more errors will be appeared.

Some examples:

abnormalitie

deep liraning

resuled in accuracy

lose va;uable

computational processess

comparision with

clinical mammography with biomarker are continuously

promsing classifier’s performance

CNN can be determine

and many more. A scientific paper should be prepared with much more care and precision.

The authors should keep in mind that scrutinizing a manuscript, or put it in order, is NOT the reviewers’ task. Alone the high number of errors enumerated by the reviewers show clearly and unambiguously that the manuscript was (and still it is) not ripe for publication. Obviously, the authors lack the necessary scientific school. They have definitely misunderstood their role in the reviewing process. The referees work for free and try to help you to produce a better, more readable manuscript. The reviewers do not need extended explanations; commonly, they know the issue better than the authors themselves. Alternatively, the manuscript should be adjusted to achieve the highest scientific standards. The response should be brief and concise and to the point.

Response: Thank you for reminding us. Some experimental results and some sentences have been added in Sections 3.3, in Pages#11, #13, and #14, and some words have been corrected.

Author Response File: Author Response.pdf

Article Menu

Printed Edition

Automatic Breast Tumor Screening of Mammographic Images with Optimal Convolutional Neural Network

Further Information

Guidelines

MDPI Initiatives

Follow MDPI