Deep Learning Approaches with Digital Mammography for Evaluating Breast Cancer Risk, a Narrative Review

Round 1

Reviewer 1 Report

This is a literature review and description of current AI and DL techniques in breast cancer risk modeling in the purpose of optimizing BC screening recommendations in the individual level. The topic is important and of clinical importance therefore clinically interesting.

The article is a well written and generally I do not have major comments.

I suggest adding in the title that this is a review of literature rather than an original study.

In some areas along the article the authors should add references, such as papers presented in Table 1.

I also suggest adding a conclusion sentence to wrap up the article.

Author Response

I suggest adding in the title that this is a review of literature rather than an original study.

New title - Advancements in Breast Cancer Risk Assessment: Review of AI and Deep Learning Approaches with Digital Mammography

In some areas along the article the authors should add references, such as papers presented in Table 1.

Added references at 3.86

Added references at 4.133 4.137

Added reference at Table 1 and altered the formatting in Table 1 to show the 7 datasets are subset of the Yala article.

I also suggest adding a conclusion sentence to wrap up the article.

Overall, AI models outperform the clinical models currently in use, though these AI tools are in the validation stages of development. With concordance indices averaging 0.78 [22], there is still much room for improvement and many new methods to apply. Application of these models could help fine tune screening practices beyond traditional risk factors that apply to broad populations and qualitative imaging characteristics.

Reviewer 2 Report

The article "Advancements in Breast Cancer Risk Assessment: AI and Deep Learning Approaches utilizing Digital Mammography" provides an insightful overview of the use of artificial intelligence (AI) and deep learning techniques in breast cancer risk modeling. The authors emphasize the need for personalized risk models and the potential for AI methods to improve the accuracy of breast cancer risk assessment.

One minor revision that could be made to the article is to elaborate on how the AI models compare to radiologists in terms of accuracy. It would be interesting to know how the AI models perform in comparison to human experts and how the two approaches could be combined for even more accurate risk assessment.

The authors also touch on the importance of the area under the curve (AUC) in determining the clinical viability of the AI models. It would be helpful to further emphasise what AUC value would be required for the AI models to achieve to be used in a clinical setting, as this would provide context for the potential impact of this technology.

The article should highlight the key differences between AI models that achieve relatively high performance, such as the Barretos model, and others. Perhaps the choice of the pretrained model and specific preprocessing techniques may contribute to the model's success, these variables could be added to the table 1. This information could be further elaborated on to provide insight into what specific elements of the AI models are contributing to their high performance.

One potential weakness of using accuracy as a metric with unbalanced data should also be noted in the article. This is an important point to consider, as it underscores the need for careful evaluation and interpretation of AI models.

Overall, the article provides a comprehensive overview of the potential for AI and deep learning techniques to improve breast cancer risk assessment. The minor revisions mentioned above could further enhance the clarity and impact of the article.

Author Response

The article "Advancements in Breast Cancer Risk Assessment: AI and Deep Learning Approaches utilizing Digital Mammography" provides an insightful overview of the use of artificial intelligence (AI) and deep learning techniques in breast cancer risk modeling. The authors emphasize the need for personalized risk models and the potential for AI methods to improve the accuracy of breast cancer risk assessment.

One minor revision that could be made to the article is to elaborate on how the AI models compare to radiologists in terms of accuracy. It would be interesting to know how the AI models perform in comparison to human experts and how the two approaches could be combined for even more accurate risk assessment.

Added to section 3: Radiologists use the standardized BI-RADS system to evaluate mammograms and document imaging features. Some features, such as breast density, have been correlated with breast cancer risk. However, radiologists do not directly predict a patient's risk of developing breast cancer.

Added to section 2: Increasingly, clinicians integrate imaging findings with current clinical models to evaluate a patient's cancer risk.

The authors also touch on the importance of the area under the curve (AUC) in determining the clinical viability of the AI models. It would be helpful to further emphasise what AUC value would be required for the AI models to achieve to be used in a clinical setting, as this would provide context for the potential impact of this technology.

We add to 5.181-186:

Overall, AI models outperform the clinical models currently in use, though these AI tools are in the validation stages of development. With concordance indices averaging 0.78 [22], where an AUC of 0.7 or higher is generally considered acceptable for a risk prediction model to be useful, there is still much room for improvement and many new datasets to train with. Application of these models could help fine tune screening practices beyond traditional risk factors, which ap-ply to broad populations and qualitative imaging characteristics.

The article should highlight the key differences between AI models that achieve relatively high performance, such as the Barretos model, and others. Perhaps the choice of the pretrained model and specific preprocessing techniques may contribute to the model's success, these variables could be added to the table 1. This information could be further elaborated on to provide insight into what specific elements of the AI models are contributing to their high performance.

Added references to Table 1 and reformatting to show the 7 datasets are subset of the Yala article. Added a methods column to highlight the use of risk factors

One potential weakness of using accuracy as a metric with unbalanced data should also be noted in the article. This is an important point to consider, as it underscores the need for careful evaluation and interpretation of AI models.

                Highlighted the importance of auc with insertion on pg 3 line 93-96

Area Under the Curve (AUC) of the receiver operator characteristic (ROC) curve is used to evaluate diagnostic performance, especially important given the overwhelming prevalence cancer negative on screening mammograms.

                Highlighted the importance of c-index with insertion on pg 4 line 126-129:

A concordance index (c-index), the generalization of AUC, shows the weighted performance of the predicted 1-5 year cancer risk.           

Overall, the article provides a comprehensive overview of the potential for AI and deep learning techniques to improve breast cancer risk assessment. The minor revisions mentioned above could further enhance the clarity and impact of the article.

Thank you