Brain Tumor Recognition Using Artificial Intelligence Neural-Networks (BRAIN): A Cost-Effective Clean-Energy Platform

Round 1

Reviewer 1 Report

Glioblastoma Multiforme or Grade 4 Astrocytoma is a fatal brain tumor with 99% mortalilty rate, however, early detection could increase the survival rate by effective surgery and new generation monoclonal antibody therapy; MRI and PET scans are the best way to diagnose, however, it is a challenge for radiologists to effectively identify and give prognosis of the cancer; The authors in this manuscript have developed a new deep learning based model system employing AI and Machine learning . By leveraging automated optimized learning, they developed a cost-effective DL platform that accurately classified brain tumors from axial MRI images of different levels.

The authors have developed an improved CNN model that can accurately classify brain tumors in  brain MR images for the purpose of diagnosing brain cancer. Experimental results  showed that the proposed model performed robustly and can be easily applied in  healthcare systems. This CNN model offers several advantages over existing approaches,  including a low carbon footprint, ease of use, and resource optimization; I strongly recommend acceptance of this manuscript for publication as this paper will definitely help the Radiologist community to effectively diagnose and predictive prognosis.

Author Response

We thank reviewer 1 for acknowledging the development of our deep learning platform that is cost-effective and clean energy.  The reviewer has recommended acceptance of this manuscript without revisions. 

Reviewer 2 Report

This review deals with the important issue of diagnosis of brain tumors, however; some points should be clarified:

1.     It is not clear whether the authors describe a review or pretend to describe an original finding (a platform; P1,L15).

2.     The title does not indicate the use of AI, nor how a cost-effective clean-energy (?) platform could be used.

3.     The intended platform is not described, the authors mention the use of 3 databases (P3,p1,L104-114) but do not explain the findings related to the possible combination of data nor the characteristics of the diagnostic technique.

4.     If this is not a systematic review, but the synthesis of an original investigation, the lack of contents and figures related to specificity of the “platform” should be described, this is important due to the high sensitivity found (96.8%).

5.     If this study is presented as an original investigation, it should be mentioned that the “platform” was made not by AI but by digital diagnostic “imaging” therefore it is cost-effective, rather than due to clean-energy (carbon-print of 0.0014 kg).

6. The number of patients included in this study must be mentioned together with the number of images (2611; P3,p1,L109).

No additional comments.

Author Response

• It is not clear whether the authors describe a review or pretend to describe an original finding (a platform; P1,L15).
  • We apologize if our methodology was not clear. In this manuscript we first performed a systematic review of the literature to define the current landscape of applying CNN for brain tumor classification. Then, we presented our  cloud-based platform.
• The title does not indicate the use of AI, nor how a cost-effective clean-energy (?) platform could be used.
  • Thank you for this comment. We just wanted to clarify that convolutional Neural Networks (CNNs) is a component of artificial  intelligence which is why we included it in the title. All of the components that we used is based on artificial intelligence tools. We also demonstrated the applications of this platform by validating our results on different images from those that were used for model training. 
• The intended platform is not described, the authors mention the use of 3 databases (P3,p1,L104-114) but do not explain the findings related to the possible combination of data nor the characteristics of the diagnostic technique.
  • We specified "To further validate the generalizability of our model’s performance, we conducted further testing on a separate cohort (100 naïve images from each dataset distinct from our training, validation, and test data that has not been seen by our algorithm).
  • The creators of this database confirmed the diagnosis of all the images used. We specify how our platform could identify the diagnoses of MRI images purely on the features on the images to similar accuracy as the confirmed radiological and pathological assessments. The diagnostic technique that the platform performs is unknown and was not the focus of this study. Such information can be obtained by subsequent analyses looking at the AI explainability metrics to figure out which feature parameters were useful to the machine to coming up with the brain tumor diagnosis.
• If this is not a systematic review, but the synthesis of an original investigation, the lack of contents and figures related to specificity of the “platform” should be described, this is important due to the high sensitivity found (96.8%).
  • The authors are unsure what is meant by the reviewer. Does the reviewer suggest we include images of the code used to train and validate our algorithm/platform? The authors thought that such information is outside of the scope of this manuscript and Journal. We added some clarifying information about how the platform was built using Google Collab a cloud-based notebook service that allowed us to develop and test our code, then utilize free super-computing resources to train and deploy our model (P4, p1, L120-124) (see revision 1 attachment). Please advise which specific figure/content is recommended by the reviewer.
• If this study is presented as an original investigation, it should be mentioned that the “platform” was made not by AI but by digital diagnostic “imaging” therefore it is cost-effective, rather than due to clean-energy (carbon-print of 0.0014 kg).
  • The platform was made by the authors by utilizing Artificial intelligence tools to perform the classification automatically on digital diagnostic imaging (ie. MRI). There was no manual input needed after running the algorithm since the our machine learned the parameters from the training dataset that we used. The point of our manuscript was to develop a platform that can train a model to learn features from images that we know the brain tumor diagnosis, then apply this knowledge to naive images to predict the brain tumor diagnosis. 
• The number of patients included in this study must be mentioned together with the number of images (2611; P3,p1,L109). 
  • Thank you for this point. Although, we understand the reviewer's comment, unfortunately the datasets that were used did not specify the number of patients to adhere to patient privacy as per the dataset creators. Other published studies have used similar datasets with the same limitations.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The manuscript has been improved.

No comments.