Performance of Machine Learning and Texture Analysis for Predicting Response to Neoadjuvant Chemoradiotherapy in Locally Advanced Rectal Cancer with 3T MRI
Round 1
Reviewer 1 Report
The authors report encouraging results in the prediction of Pathological Complete Response (pCR) to neoadjuvant Chemoradiotherapy (nCRT) for Locally Advanced Rectal Cancer Patients analysing Texture parameters, including skewness, kurtosis, entropy and mean of positive pixels extracted from staging 3T pelvic MRI. A Machine Learning-based approach demonstrated useful to combine all Texture Analysis parameters and filers to obtain a diagnostic model (decisional tree) with the ability to discriminate a priori pCR patients after nCRT, with the clinical implications in the modern management of rectal cancer patients as well reported by the authors.
In spite of some limitations as already reported by the authors, this is a well designed and well described study, and of particular interest in the perspctive of treatment personalization for patients with rectal cancer
Further comments:
- This is a prospective study and I suggest to emphasised that in the discussion as many radiomic/machine learning reported studies are retrospective
- The restaging (post-treatment) MRI has done at the end of nCRT (page 3, line 106 and fig1) resulting in early evaluation (clinical Response) compared to time of surgery (6-8wks) with path Response; this could be a bias in clinical and path Response correlation (data in time/response are well defined in rectal cancer). This issue should be reported and discussed in the discussion
- Did the authors comply to the Image Biomarker Standardisation Initiative (IBSI)? This would be very important for the standardisation of the whole process. In any case, also this should be mention in the discussion
-Did the authors extract also textural features (e.g. GLCM, NTDM etc) in addition to that reported in the text, mostly first order features (entropy, skewness etc.). If textural features were not collected, we encourage the authors to do so as with Textrad it should be feasible (Texture Analysis is the main issue of the paper)
- In the manuscript is not reported any testing or validation of the results from machine learning: cross validation, holdout, booststrap or other. We expected the authors to perform some type of model testing or validation, otherwise the apparently nice results could represent just overfitting of data from machine learning. It is possible that Weka outputs results over cross validation, if so, you can describe this ( we suggest as example to see and citing https://www.frontiersin.org/articles/10.3389/fonc.2020.00490/full )
Minor comment
Page 11, Fig 6 change form .. with from
-
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 2 Report
Dear Authors,
Thank you very much for the opportunity to revise this manuscript: the research topic is extremely interesting and valuable, and represents one of the best applications of radiomics aimed toward clinical practice.
I would like to praise the effort made to highlight the individual features' predictive value, which is rarely seen in this kind of publications.
I also appreciate the strategy to segment a single slice of a FSE sequence, which should give the best contrast to perform TA on.
I would suggest some minor revisions to further improve the quality of the manuscript:
AIMS
This is completely optional: it would be really interesting to see the correlation of radiomic features with the tumor grade. It would also be interesting to include the tumor grade in the ML algorithm, as eyeballing the date it seems that NR/PR have a higher tumor grade than CR. Again, this is not to be considered a revision, but would increase the quality of the paper.
MATERIALS AND METHODS
- Weka should be cited as specified on their website: Eibe Frank, Mark A. Hall, and Ian H. Witten (2016). The WEKA Workbench. Online Appendix for "Data Mining: Practical Machine Learning Tools and Techniques", Morgan Kaufmann, Fourth Edition, 2016.
- I personally deem image normalization not necessary in this research setting, as the MRI protocol is standardized enough for this kind of exam. However, TA on MRI is very delicate process, as feature could vary depending on the FOV and the TE/TR. Consider referencing to this aspect.
RESULTS
- The performance of the ML model is not stated clearly. Consider adding the predicted group, group probability, and test it with a ROC analysis on the subgroup analysis you have mentioned in the discussion. Moreover, an "internal" validation could be performed by training the model on 70% of the patients and testing in on the remaining 30% (this is optional, take it as a suggestion and not a revision).
-consider performing statistical tests to assess differences and similarities among the two/three groups (CR vs PR vs NR) as far as the demographics variables are concerned in table 1.
DISCUSSION
- you state that you did not compare TA with other imaging biomarkers; could you provide some examples of other known biomarkers (for example dwi/adc)?
-similar results to those you have obtained have also been achieved by the group of Coppola F et al; try to discuss similarities and differences with other papers ( I am suggesting this for reference, feel free to integrate whichever you wish Coppola F, Mottola M, Lo Monaco S, Cattabriga A, Cocozza MA, Yuan JC, De Benedittis C, Cuicchi D, Guido A, Rojas Llimpe FL, D'Errico A, Ardizzoni A, Poggioli G, Strigari L, Morganti AG, Bazzoli F, Ricciardiello L, Golfieri R, Bevilacqua A. The Heterogeneity of Skewness in T2W-Based Radiomics Predicts the Response to Neoadjuvant Chemoradiotherapy in Locally Advanced Rectal Cancer. Diagnostics (Basel). 2021 Apr 28;11(5):795. doi: 10.3390/diagnostics11050795. PMID: 33924854; PMCID: PMC8146691.)
Author Response
Please see the attachment.
Author Response File: Author Response.docx