MRI-Based Radiomics for Non-Invasive Prediction of Molecular Biomarkers in Gliomas

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The submitted review is a systemic review of radiomic analysis of MRI to predict molecular biomarkers in gliomas. The topic should be of interest to the readership.

The strengths of the paper include a good number of studies (70 or 71) and a reasonable number of patients (10,324) in the review. The prediction of multiple biomarkers of interest using radiomics is evaluated. The review also includes a good review of quality of the included studies including assessments based on RQS, IBSI, and NOS. Multiple methodologies are also reviewed, including methods based on human-engineered features and others based on deep learning.

However, it is indicated that "a focused and systematic evaluation dedicated specifically to the integration of radiomics and molecular biology in gliomas is currently absent." However, there are a few, including those in the cited references:

Di Salle G, Tumminello L, Laino ME, Shalaby S, Aghakhanyan G, Fanni SC, et al. Accuracy of Radiomics in Predicting IDH Mutation Status in Diffuse Gliomas: A Bivariate Meta-Analysis. Radiology: Artificial Intelligence [Internet]. 2023 Dec 6 [cited 2025 Dec 30]; Available from: https://pubs.rsna.org/doi/10.1148/ryai.220257

Ahmadzadeh AM, Lomer NB, Ashoobi MA, Bathla G, Sotoudeh H. MRI-derived radiomics and end-to-end deep learning models for predicting glioma ATRX status: a systematic review and meta-analysis of diagnostic test accuracy studies. Clinical Imaging [Internet]. 2025 Mar 1 [cited 2025 Dec 30];119:110386. Available from: https://www.sciencedirect.com/science/article/pii/S0899707124003164

Others exist, including:

Chung CYC, Pigott LE. Predicting IDH and ATRX mutations in gliomas from radiomic features with machine learning: a systematic review and meta-analysis. Front Radiol. 2024 Oct 31;4:1493824. doi: 10.3389/fradi.2024.1493824. PMID: 39544481; PMCID: PMC11560782.

Ahmadzadeh AM, Broomand Lomer N, Ashoobi MA, Elyassirad D, Gheiji B, Vatanparast M, et al. MRI-derived deep learning models for predicting 1p/19q codeletion status in glioma patients: a systematic review and meta-analysis of diagnostic test accuracy studies. Neuroradiology [Internet]. 2025 July 1 [cited 2025 Dec 30];67(7):1667–81. Available from: https://doi.org/10.1007/s00234-025-03631-z

Farahani S, Hejazi M, Tabassum M, Di Ieva A, Mahdavifar N, Liu S. Diagnostic performance of deep learning for predicting glioma isocitrate dehydrogenase and 1p/19q co-deletion in MRI: a systematic review and meta-analysis. Eur Radiol [Internet]. 2025 Aug 16 [cited 2025 Dec 30]; Available from: https://doi.org/10.1007/s00330-025-11898-2

These are quite recent and the submitted paper indicates the authors performed their search of online reference databases on 2/1/2025. However, since these other references are available now, they should probably be referenced. I only searched briefly for other reviews, so a more thorough search is likely warranted. The submitted review has strengths over these including the larger number of molecular markers, the thorough assessment of quality of the included studies, and the multiple methodologies of the included studies, as alluded to when listing the strengths of the paper.

Although the authors provide assessments of quality of the included studies based on RQS, IBSI, and NOS, there is not much discussion about the results. It is interesting that many seem to be of high quality based on NOS, but most have relatively low scores as assessed by RQS and many are low to mediocre based on IBSI. Some discussion of these incongruent results would be interesting.

The paper provides a good overview of the different aspects of radiomics as would be expected in a systemic review.  There are some other points that might be interesting to discuss further. For example, clinically we almost always use diffusion imaging, but only a few of the analyzed studies seem to include it although it may provide information about underlying tissue structure. Discussion of possible reasons (e.g. difficulty with the image processing/analysis, artifacts, and/or weaker result) would be interesting. There is some discussion of MR perfusion and spectroscopy, which is appreciated.

It is mentioned that meta-analysis was not performed because of heterogeneity. However, it seems that there have been meta-analyses that have been performed in other reviews. More discussion about this would be appreciated.

A few other minor errors or inconsistencies are present:

1. The stated number of studies reviewed is inconsistent and requires clarification. Figure 3 indicates 71 studies, while the text states 70 studies in the abstract and discussion and 74 studies under results.
2. Near the bottom of page 8, the summary of included studies appears to be in Table 1 rather than the stated Table 3.

Author Response

1 - These are quite recent and the submitted paper indicates the authors performed their search of online reference databases on 2/1/2025. However, since these other references are available now, they should probably be referenced.

R1 - We thank the reviewer for pointing out these recent publications. We have now incorporated these references into the revised manuscript to ensure the literature review is as up-to-date as possible.

 

2 - Although the authors provide assessments of quality of the included studies based on RQS, IBSI, and NOS, there is not much discussion about the results. It is interesting that many seem to be of high quality based on NOS, but most have relatively low scores as assessed by RQS and many are low to mediocre based on IBSI. Some discussion of these incongruent results would be interesting.

R2 - We thank the reviewer for this insightful observation. We have added a paragraph in the Discussion (4.5) to address the discrepancy between the quality scores. We note that while the studies show high clinical quality, they often lack the technical standardization and reporting transparency required by radiomics-specific tools like RQS and IBSI. This gap highlights the need for better standardization in the field.

 

3 - For example, clinically we almost always use diffusion imaging, but only a few of the analyzed studies seem to include it although it may provide information about underlying tissue structure. Discussion of possible reasons (e.g. difficulty with the image processing/analysis, artifacts, and/or weaker result) would be interesting.

R3 - We thank the reviewer for this valuable suggestion. We have updated the Discussion section to address the limited use of DWI in the reviewed literature.

 

4 - It is mentioned that meta-analysis was not performed because of heterogeneity. However, it seems that there have been meta-analyses that have been performed in other reviews. More discussion about this would be appreciated.

R4 - We thank the reviewer for this comment. We have expanded the Discussion to clarify our decision. Although other reviews have performed meta-analyses, the studies we analyzed were extremely diverse in terms of technology and clinical endpoints. We believe that pooling such heterogeneous data would lead to unreliable results. Therefore, we focused on a qualitative synthesis to provide a more accurate and rigorous overview of the current literature.

 

5 - The stated number of studies reviewed is inconsistent and requires clarification. Figure 3 indicates 71 studies, while the text states 70 studies in the abstract and discussion and 74 studies under results.

R5 - We thank the reviewer for identifying these inconsistencies. We have carefully cross-checked the data and corrected the discrepancy throughout the manuscript, abstract, and figures. The correct number of included studies is 70, and all sections have been updated to ensure consistency.

 

6 - Near the bottom of page 8, the summary of included studies appears to be in Table 1 rather than the stated Table 3.

R6 - Thank you for the correction. We have fixed the error, and the text now correctly refers to Table 1.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This review paper addresses MRI-based, non-invasive prediction of molecular biomarkers in gliomas, including IDH, ATRX, MGMT, 1p/19q codeletion, etc. The issue addressed here is very important when we try to think about the therapeutic strategies for gliomas in the future. I do not mean to delve into the details of this paper/issue, which, I believe, the authors are the most familiar with, but I want to raise the following points about the article.

 

#Table 1

-second row: DTI >>> DWI (Highly probably, this must be DWI)

-Which is intended for use in this MS, 1p19q or 1p/19q? An identical thing should be identical in its style (This should be applied, not only to here but also to the whole of the MS).

-Given the above mistyping, I would like to recommend the authors to check the correctness of this table again.

-Table 1 or somewhere in the text: To assess the performance of MRI-based, non-invasive prediction of molecular biomarkers, the AUC must be very important. This is kind of a famous index in medical science, and I know that AUC ranges between 0 and 1, and a higher value is considered better. But how can we interpret this index in general? For example, could 0.3 be bad enough so that it could prevent its application to clinical medicine? Or, could 0.7 be good enough to apply it to clinical medicine? This kind of information must be very helpful for those readers not necessarily familiar with this kind of methodology to interpret the Table 1 or the whole of the paper. Not a strict one, which I guess is impossible, but something of a hint to its interpretation would be appreciated.

 

#It says in line 302, which is just in front of Table 1, ‘Table 3’ instead of Table 1. I guess this should be Table 1, right?

 

#line 317: What is dMRI? Diffusion MRI???

 

#This is similar to the above. Table 2 and 3 are okay, but why RQS, IBSI, or NOS should be specifically addressed in this paper? RQS, IBSI, or NOS is something of an index about the reliability of a paper given? Not all the readers are familiar with this kind of methodology; so, please draft the MS so that they can naturally understand what they are or why they have to be in this MS, specifically being showcased in the tables. Since a review paper is intended for unfamiliar as well as familiar readers, I want this kind of care and generosity from the authors.

 

#Discussion

-line 423: 4.1.21. p/19q >>> 4.1.2. 1p/19q (This is a minor, but a critical error)

-In the beginning part of this Discussion, the authors state ‘offering the potential to complement, or even partially replace, biopsy-based tissue characterization’. I think this may true. But I just wonder what the role of those biopsy-based tissue characterization would be like in the future. I believe it will not disappear, although with a less performance rate, even in the future. I want a subsection (4.7.?) that is specifically devoted to the discussion of biopsy-based tissue characterization/tissue biopsy in terms of its relation to the presently addressed MRI-based methodology. If I am allowed to state personally, as the authors say, ‘complement’, would be the right image expected to see in the future. Anyway, the authors’ own is okay, but I want to know what or how they think/expect/hope.

 

#Generally, this MS is well-written and nicely structured, but as is also described above, some grammatically/stylistically questionable points are noted. They may seem minor ones, but some of which are crucially unfavorable for this review article. The whole of the MS should be proofread again, with care and devotion.

Author Response

1 - Second row: DTI >>> DWI (Highly probably, this must be DWI).

R1 - We thank the reviewer for the observation. We have corrected the error in the manuscript.

 

2 - Which is intended for use in this MS, 1p19q or 1p/19q? An identical thing should be identical in its style

R2 - We thank the reviewer for the observation. We have corrected the error in the manuscript.

 

3 - I would like to recommend the authors to check the correctness of this table again.

R3 - Thank you for the observation. We have revised the tables accordingly.

 

4 - Table 1 or somewhere in the text: To assess the performance of MRI-based, non-invasive prediction of molecular biomarkers, the AUC must be very important. This is kind of a famous index in medical science, and I know that AUC ranges between 0 and 1, and a higher value is considered better. But how can we interpret this index in general? For example, could 0.3 be bad enough so that it could prevent its application to clinical medicine? Or, could 0.7 be good enough to apply it to clinical medicine? This kind of information must be very helpful for those readers not necessarily familiar with this kind of methodology to interpret the Table 1 or the whole of the paper. Not a strict one, which I guess is impossible, but something of a hint to its interpretation would be appreciated.

R4 - Thank you for this suggestion. We have added a brief guide to help readers interpret AUC values (MATERIALS AND METHODS - paragraph 2.2)

 

5 - It says in line 302, which is just in front of Table 1, ‘Table 3’ instead of Table 1. I guess this should be Table 1, right?

R5 - We appreciate the feedback. The table reference has been corrected accordingly.

 

6 - line 317: What is dMRI? Diffusion MRI?

R6 - We appreciate the feedback. The error has been corrected both in the text and within the table.

 

7 - This is similar to the above. Table 2 and 3 are okay, but why RQS, IBSI, or NOS should be specifically addressed in this paper? RQS, IBSI, or NOS is something of an index about the reliability of a paper given? Not all the readers are familiar with this kind of methodology; so, please draft the MS so that they can naturally understand what they are or why they have to be in this MS, specifically being showcased in the tables. Since a review paper is intended for unfamiliar as well as familiar readers, I want this kind of care and generosity from the authors.

R7 - Thank you for the suggestion. We have revised Section 2.4 to provide a better explanation of the RQS, IBSI, and NOS methodologies, ensuring they are easily understandable for all readers.

 

8 - Line 423: 4.1.21. p/19q >>> 4.1.2. 1p/19q (This is a minor, but a critical error)

R8 - Thank you for the feedback. The error has been fixed in the manuscript.

 

9 - In the beginning part of this Discussion, the authors state ‘offering the potential to complement, or even partially replace, biopsy-based tissue characterization’. I think this may true. But I just wonder what the role of those biopsy-based tissue characterization would be like in the future. I believe it will not disappear, although with a less performance rate, even in the future. I want a subsection (4.7.?) that is specifically devoted to the discussion of biopsy-based tissue characterization/tissue biopsy in terms of its relation to the presently addressed MRI-based methodology. If I am allowed to state personally, as the authors say, ‘complement’, would be the right image expected to see in the future. Anyway, the authors’ own is okay, but I want to know what or how they think/expect/hope.

R9 - Thank you for this suggestion. We agree that radiomics and biopsy should be seen as complementary. We have added subsection 4.7 to discuss how these two methods can work together in the future to improve patient care.

 

10 - Generally, this MS is well-written and nicely structured, but as is also described above, some grammatically/stylistically questionable points are noted. They may seem minor ones, but some of which are crucially unfavorable for this review article. The whole of the MS should be proofread again, with care and devotion.

R10 - We thank the Reviewer for the comments. The full text has been revised and updated accordingly.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

In this study, the authors systematically review radiomics and artificial intelligence-based approaches for the non-invasive prediction of molecular biomarkers in gliomas. Analyzing over 70 studies in accordance with PRISMA guidelines, the article comprehensively compares the MRI protocols, segmentation strategies, ML/DL models, and molecular targets used. Furthermore, methodological quality and clinical applicability were assessed using the RQS, IBSI, and NOS frameworks. The findings indicate that IDH mutation was the most consistently predictable biomarker, while methodological heterogeneity and insufficient external validation posed significant limitations for other biomarkers. However, I kindly offer the following suggestions to the authors:

1- A quantitative meta-analysis was not performed in this study due to heterogeneity; however, subgroup analyses or limited effect size assessments should have been performed.

2- The distinction between radiomics and DL is superficial; the differences between handcrafted radiomics and DL-based feature learning have not been analyzed in depth.

3- RQS and NOS scores are reported, but critical commentary is weak; there is insufficient questioning of why studies labeled “high quality” fail in clinical practice.

4- The topic of external validation is repeated, but no solution is proposed; the problem is correctly identified, but no methodological roadmap is presented.

5- I believe the integration of multi-omics is exaggerated. While there are very few genuine multi-omics studies in the literature, this approach, presented as a promising main direction, should be presented with a more flexible stance or with evidence to support it.

6- Tables are excessively long, and the synthesis is weak. For example, Table 1 and the NOS tables are very detailed but lack an interpretive summary. The answer to the question “What should I learn?” is not clear to the reader.

7- The Introduction, Discussion, and Conclusion sections repeat nearly identical statements. These sections should be thoroughly reviewed and revised.

8- Figures are descriptive but not analytical; PRISMA and workflow figures are standard, and there is no new conceptual framework diagram.

9- NOS and RQS are high, but the results are weak in clinical practice. This contradiction is not sufficiently explained theoretically. Similarly, DL performance is said to be high, but the discussion on bias is insufficient; overfitting, data leakage, and class imbalance are not adequately addressed.

Author Response

1 - A quantitative meta-analysis was not performed in this study due to heterogeneity; however, subgroup analyses or limited effect size assessments should have been performed.

R1 - We appreciate the reviewer’s valuable observation regarding the current state of the literature. While primary research has expanded, a focused and systematic evaluation specifically dedicated to the integration of radiomics and molecular biology in gliomas remains limited to a few specific works. To address this, we have included Table 2 to provide a more comprehensive and detailed analysis of the available results, highlighting the distribution of biomarkers, modeling strategies, and performance metrics across the included studies.

 

2- The distinction between radiomics and DL is superficial; the differences between handcrafted radiomics and DL-based feature learning have not been analyzed in depth.

R2 - We thank the reviewer for this observation; accordingly, we have incorporated Table 3 to better elucidate these aspects and provide a more detailed comparison.

 

3- RQS and NOS scores are reported, but critical commentary is weak; there is insufficient questioning of why studies labeled “high quality” fail in clinical practice.

R3 - We appreciate the reviewer’s comment. Rather than adding new content, we have refined and clarified the discussion in section 4.6 (Current challenges and future perspectives) to better address this point. Specifically, we have expanded our critical analysis of the 'translational gap' to explain why studies with high RQS/NOS scores may still encounter difficulties in clinical implementation.

 

4- The topic of external validation is repeated, but no solution is proposed; the problem is correctly identified, but no methodological roadmap is presented.

R4 - We thank the reviewer for the suggestion. We have refined Section 4.5 to describe three operative solutions—data harmonization, federated learning, and prospective trials.

 

5 - I believe the integration of multi-omics is exaggerated. While there are very few genuine multi-omics studies in the literature, this approach, presented as a promising main direction, should be presented with a more flexible stance or with evidence to support it.

R5 - We thank the Reviewer for the feedback. We agree that multi-omics research is still in its early stages. We have revised the text to be more cautious and realistic, describing it as a 'potential' rather than a 'major' frontier.

 

6 - Tables are excessively long, and the synthesis is weak. For example, Table 1 and the NOS tables are very detailed but lack an interpretive summary. The answer to the question “What should I learn?” is not clear to the reader.

R6 - Regarding the presentation of the data, we have elected to maintain the comprehensive nature of the tables to ensure maximum methodological transparency and to provide a complete overview of the literature’s evolution. To facilitate readability without losing critical information, we have included detailed descriptions of the specific objectives and findings for each table directly within their captions.

 

7 - The Introduction, Discussion, and Conclusion sections repeat nearly identical statements. These sections should be thoroughly reviewed and revised.

R7 - We thank the Reviewer for this important observation. We have thoroughly reviewed the Introduction and Discussion sections to remove redundant statements and ensure a more distinct and logical flow.

 

8 - Figures are descriptive but not analytical; PRISMA and workflow figures are standard, and there is no new conceptual framework diagram.

R8 - We thank the Reviewer for the feedback. To ensure maximum transparency, the original descriptive tables have not been modified. However, we have added Tables 2, 3, and 6 to provide a better interpretation of the results and to address the specific points raised during the revision.

 

9 - NOS and RQS are high, but the results are weak in clinical practice. This contradiction is not sufficiently explained theoretically. Similarly, DL performance is said to be high, but the discussion on bias is insufficient; overfitting, data leakage, and class imbalance are not adequately addressed.

R9 - We thank the Reviewer for this profound observation. We have added a dedicated paragraph in the Discussion to address the 'quality-to-practice' gap. We now clarify that high RQS and NOS scores reflect methodological rigor but do not necessarily guarantee clinical utility (DISCUSSION - 4.5). Additionally, we have expanded the discussion on Deep Learning limitations, specifically addressing how overfitting, data leakage, and class imbalance can bias performance results and hinder their real-world application.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The revision is satisfactory. The paper will be read with interest by many readers. 

Reviewer 3 Report

Comments and Suggestions for Authors

I would like to thank the authors for their efforts; they have made the suggested corrections. In this regard, I kindly recommend that the study be accepted.