Radiomic Analysis for Ki-67 Classification in Small Bowel Neuroendocrine Tumors

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

 

 Radiomic Analysis for Ki-67 Classification in Small Bowel Neuroendocrine Tumors

The manuscript is clearly written, specific, and reflects the study’s focus on radiomics and Ki‑67 classification in SB‑NETs. A few comments are mentioned below:

1. The introduction section should include a more detailed discussion of prior SB‑NET radiomics studies.
2. The manuscript should be checked for typographical errors. Like for example, line nos 100-101, “We initially selected 54 patients from two different institutes with pathology proven small bowel neuroendocrine tumor, ho underwent at least a pre-treatment contrast-media CT exam,-----.
3. Though a rare condition, the dataset size of 34 patients with 128 lesions is small, limiting generalizability.
4. How were the hyperparameters tuned for each ML model.
5. Variability in XGBoost performance is mentioned within the manuscript; however, a clear explanation is missing.
6. Add details of all the machines or instruments used for this study in the methodology section.
7. The limitation section could be expanded, specifically highlighting the retrospective design and heterogeneity of CT acquisition protocols, among other factors.
8. I could not find any external validation mentioned within the manuscript.

Author Response

Comment 1: The introduction section should include a more detailed discussion of prior SB‑NET radiomics studies.

Thanks for pointing this out. In the introduction we pointed out that the vast majority of GEP NET studies focused on pancreatic NETs, and we cited that only one study analyzed patients with ileal neuroendocrine tumors but without focusing on grading, while predicting the risk of developing complications due to mesenteric masses using clinical criteria and radiomics.

I added the numbers showing the growth of interest for radiomics applied to NEts in recent years.

Comment 2: The manuscript should be checked for typographical errors.

I'm sorry our English was not 100% on point. We carefully checked every paragraph, ensuring there are no typographical errors in this version.  

Comment 3. Though a rare condition, the dataset size of 34 patients with 128 lesions is small, limiting generalizability.

Thanks for the comment, you are absolutely right. To implement the dataset of 34 patients we evaluated every lesion visualized on CECT scan (primary tumor, pathological lymph nodes and liver metastases) and considered each segmented lesion as an independent lesion, for a total of 128 lesions. Nevertheless, we know the number criticity remains and we underlined it in the study limits. 

Comment 4. How were the hyperparameters tuned for each ML model. 

Thanks for the comment. Hyperparameter optimization was performed using GridSearchCV within the training set of each fold during the cross-validation process, in order to avoid data leakage. For each machine learning algorithm a predefined grid of model-specific hyperparameters was explored, and the optimal configuration was selected based on classification accuracy. Final model performance metrics were obtained by averaging the results across the five cross-validation folds.

Comment 5: Variability in XGBoost performance is mentioned within the manuscript; however, a clear explanation is missing.

The variability observed in XGBoost performance is likely related to the limited sample size and the high dimensionality of the radiomic feature space. In small datasets, boosting-based tree models may be more sensitive to data partitioning and parameter selection, resulting in fluctuations across cross-validation folds. Although synthetic oversampling techniques were applied exclusively to the training set, their use may have contributed to increased variability in ensemble boosting methods.

Comment 6: Add details of all the machines or instruments used for this study in the methodology section. 

Thanks for the comment. I will add details of the segmentation program ITK- SNAP and details regarding the statistical analysis we did

Comment 7. The limitation section could be expanded, specifically highlighting the retrospective design and heterogeneity of CT acquisition protocols, among other factors.

You are right. I added that in the manuscript limitations paragraph. 

Comment 8. I could not find any external validation mentioned within the manuscript. 

We acknowledge the reviewer's comment regarding external validation. Due to the limited number of patients and lesions in this rare tumor population, only internal cross-validation was possible at this stage. Despite this limitation, the Random Forest model demonstrated robust performance. We explicitly emphasize that external validation on independent datasets will be conducted as soon as a larger number of cases and CT images are available. This remains an important step for future research and operational implementation.

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Author

From my point of view as the molecular biologist your manuscript entitled "Radiomic Analysis for Ki-67 Classification in Small Bowel Neuroendocrine Tumors"  is excellent. The conclusion to investigate radiomic features extracted from CT images focusing on small bowel NETs and evaluate their association with Ki-67 expression is perfect because:

Ki-67 is a nuclear protein expressed in proliferating cells.

It is widely used as a prognostic and predictive biomarker in cancers (e.g., breast cancer, CNS lymphoma) and high Ki-67 levels correlate with rapid tumor growth, poor prognosis, and treatment response.

However I recommend the radiologist and a Human AI research scientists to evaluate the manuscript.

Author Response

Comment 1: Dear Author

From my point of view as the molecular biologist your manuscript entitled "Radiomic Analysis for Ki-67 Classification in Small Bowel Neuroendocrine Tumors"  is excellent. The conclusion to investigate radiomic features extracted from CT images focusing on small bowel NETs and evaluate their association with Ki-67 expression is perfect because:

Ki-67 is a nuclear protein expressed in proliferating cells.

It is widely used as a prognostic and predictive biomarker in cancers (e.g., breast cancer, CNS lymphoma) and high Ki-67 levels correlate with rapid tumor growth, poor prognosis, and treatment response.

However I recommend the radiologist and a Human AI research scientists to evaluate the manuscript.

 

Response: Thank you so much for the support. I will improve the points suggested by the remaining Reviewers!

Reviewer 3 Report

Comments and Suggestions for Authors

General remarks

In this paper the authors have evaluated the radiomics features of small bowel Neuroendocrine Tumors and their association with Ki-67 expression. This is overall a unique study. However, the number of cases are too less to have a definite comment.

Specific comments

1. The number of cases are only 34. The number is too less.
2. Currently people are using deep learning technique on radiomics. It was expected that the authors should use deep learning.
3. Why the authors took less than 1% Ki67 index, when the low grade neuroendocrine tumour is upto 2%?

Author Response

1. The number of cases are only 34. The number is too less.

Thanks for the comment, you are absolutely right. To implement the dataset of 34 patients we evaluated every lesion visualized on CECT scan (primary tumor, pathological lymph nodes and liver metastases) and considered each segmented lesion as an independent lesion, for a total of 128 lesions. Nevertheless, we know the number criticity remains and we underlined it in the study limits. 

2. Currently people are using deep learning technique on radiomics. It was expected that the authors should use deep learning. 

Thanks for the comment. As we explained in the conclusions, we did not apply deep learning in this study because of the paucity of data in our dataset. With just 128 lesions we could have not trained efficently a  neural link in a DL model, therefore there would have been too much overfitting. Instead we adopted a more appropriated ML model, which demostrated solid results in literature. 

3. Why the authors took less than 1% Ki67 index, when the low grade neuroendocrine tumour is upto 2%? Thanks for the comment. As we explained in the article, we used a more tailored cut-off of 1% applied to Ki67 index because the vast majority of SB-NETs are low grade, although many are prognostically unfavourable. Therefore there is prognostic heterogeneity in the "Ki67 < 3%" group. The cutoff of 1% for Ki67 is therefore a reliable prognostic factor.   We cited the literature we used to support this choice (REVIEW Klöppel G, La Rosa S. Ki67 labeling index: assessment and prognostic role in gastroenteropancreatic neuroendocrine neoplasms. Virchows Arch. 2018 Mar;472(3):341-349. doi: 10.1007/s00428-017-2258-0. Epub 2017 Nov 13. Erratum in: Virchows Arch. 2018 Mar;472(3):515. doi: 10.1007/s00428-017-2283-z. PMID: 29134440).

  

Reviewer 4 Report

Comments and Suggestions for Authors

This manuscript investigates whether CT-based radiomic features can classify small-bowel neuroendocrine tumor lesions by Ki-67 expression using a tailored threshold (≤1% vs >1%). The authors extract 107 PyRadiomics features from portal-phase CT with manual lesion segmentation, perform feature filtering/selection, and train multiple ML classifiers with 5-fold cross-validation. The study is clinically relevant and timely, but several methodological and reporting issues (unit-of-analysis, leakage, harmonization, and validation design) need to be addressed to support the strength and generalizability of the conclusions. 

1. The authors explicitly treat each segmented lesion (primary/nodes/liver mets) as an “independent unit” (n=128) from 34 patients. This violates independence and can inflate statistical significance and ML performance. Recommended fixes:
• Use patient-level grouping for CV (GroupKFold).
• Report patient-level performance (e.g., aggregate lesion predictions per patient via mean/max probability).
• Use cluster-robust SEs for inferential comparisons; for modeling, consider mixed-effects or at minimum grouped evaluation. 
2. The authors assume secondary lesions share primary Ki-67 because many metastases were not biopsied. That may be pragmatic but should be clearly labeled as a limitation and potential source of mislabeling. If possible, perform sensitivity analyses excluding lesions without lesion-specific pathology (or restricting to primaries/biopsied). 
3. CT acquisition differences (scanner, slice thickness, kernels, contrast timing) can dominate radiomic signals. You mention two institutes and multiple years, but don’t report acquisition parameters or how differences were handled. At minimum include a table of key acquisition variables and incorporate harmonization or stratified analysis. 
4. Clarify why you chose 1% as the threshold and how it maps to current WHO/ENETS grading; make the rationale concise and consistent. 
5. Provide median/mean lesion volume per class; performance may be driven by size/partial volume effects. 
6. Fix typos/formatting: “pat erns”, “overfit ing”, “ngtdmStrenght”, “metastases” pluralization, and duplicated “Informed Consent Statement”. 

Author Response

The authors explicitly treat each segmented lesion (primary/nodes/liver mets) as an “independent unit” (n=128) from 34 patients. This violates independence and can inflate statistical significance and ML performance. Recommended fixes:

• • Use patient-level grouping for CV (GroupKFold).
  • Report patient-level performance (e.g., aggregate lesion predictions per patient via mean/max probability).
  • Use cluster-robust SEs for inferential comparisons; for modeling, consider mixed-effects or at minimum grouped evaluation. 

We thank the Reviewer for the comment regarding potential non-independence of lesions originating from the same patient. We would like to clarify, however, that the present study was intentionally designed as a lesion-level radiomics analysis, rather than as a patient-level prognostic or decision model.

This choice was driven by specific clinical and methodological considerations:

Clinical rationale (inter-lesion heterogeneity): patients with small-bowel neuroendocrine tumors frequently present with multiple lesions (primary tumor, nodal/mesenteric disease, liver metastases), which may exhibit substantial intra-patient inter-lesion imaging heterogeneity. The aim of this study was to assess whether CT-derived radiomic patterns are associated with Ki-67 expression at the lesion scale, which represents the natural unit of radiological assessment and imaging follow-up.

Disease rarity and patient sample size: the cohort includes 34 patients, reflecting the rarity of SB-NETs. A strictly patient-level framework would have markedly reduced the available information and limited the feasibility of an exploratory analysis aimed at detecting meaningful radiomic signals. In this context, a lesion-level approach represents a pragmatic and commonly adopted strategy in radiomics studies on rare tumors.

Accordingly, the results should be interpreted as lesion-level and hypothesis-generating, in line with the study design. The purpose was not to provide definitive patient-level clinical performance estimates, but to explore the potential of radiomic features to capture imaging patterns associated with Ki-67 expression in SB-NETs.

 

The authors assume secondary lesions share primary Ki-67 because many metastases were not biopsied. That may be pragmatic but should be clearly labeled as a limitation and potential source of mislabeling. If possible, perform sensitivity analyses excluding lesions without lesion-specific pathology (or restricting to primaries/biopsied). 

Thank you for the comment, we will explore more details and will underline this limit in our discussion

 

CT acquisition differences (scanner, slice thickness, kernels, contrast timing) can dominate radiomic signals. You mention two institutes and multiple years, but don’t report acquisition parameters or how differences were handled. At minimum include a table of key acquisition variables and incorporate harmonization or stratified analysis. 

Thank you for your comment. Unfortunately we don't have all the parameters of the different CT scans in the different years of acquisition ofg the study exams (2012 - 2024), so we cannot do a table of key acquisition variables.

Clarify why you chose 1% as the threshold and how it maps to current WHO/ENETS grading; make the rationale concise and consistent. 

Thank you. As we explained we chose was chosen because the majority of SB-NETs are low-grade (up to 75% G1) [18], but a significant percentage has the tendency to develop local-regional and/or distant metastases. Therefore there is an important prognostic heterogeneity in the "Ki67 < 3%" group. The cutoff of 1% for Ki67 is consequentially a reliable prognostic factor, as highlighted in literature 

Provide median/mean lesion volume per class; performance may be driven by size/partial volume effects. 

Thank you.  I upgraded with the maximum, minimum and mean volume of all the lesions.

Fix typos/formatting: “pat erns”, “overfit ing”, “ngtdmStrenght”, “metastases” pluralization, and duplicated “Informed Consent Statement”.

Thank you for the comment; we will for sure double check the typos

Reviewer 5 Report

Comments and Suggestions for Authors

This manuscript addresses an important and relatively underexplored topic, namely the application of CT-based radiomics and machine-learning approaches for non-invasive Ki-67 stratification in small bowel neuroendocrine tumors (SB-NETs), using a clinically meaningful low-threshold cut-off. Overall, the study is well conceived, methodologically sound, and clearly positioned within the current radiomics literature. The results are promising and potentially clinically relevant.

The introduction provides a solid and coherent background on SB-NET epidemiology, prognostic relevance of Ki-67, and the rationale for radiomics. The gap in the literature is clearly identified, particularly regarding the limited focus on SB-NETs and the use of tailored Ki-67 cut-offs. Relevant and up-to-date references are included. Minor stylistic refinements could further improve clarity and conciseness, but no major conceptual additions are required.

The retrospective, multicenter design is appropriate for an exploratory radiomics study in a rare tumor entity. The lesion-based analysis, although potentially introducing within-patient correlation, is clearly explained and justified given the limited sample size. The choice of a 1% Ki-67 threshold is well motivated and clinically meaningful.

The methodology is described in substantial detail, allowing reproducibility. Image acquisition, segmentation, preprocessing, feature extraction, feature reduction, and machine-learning pipelines are all clearly reported and consistent with IBSI recommendations. The use of cross-validation and SMOTE to address class imbalance is appropriate. One point that could be clarified further is the potential impact of assigning the primary tumor Ki-67 value to metastatic lesions, which is acknowledged later as a limitation.
Results are clearly presented and logically structured. Performance metrics are appropriate and comprehensively reported. Tables and figures are generally clear and informative. The emphasis on recall and false-negative rate is clinically relevant and well justified.
The discussion adequately contextualizes the findings within existing literature, particularly in comparison with pancreatic and rectal NET radiomics studies. Strengths and limitations are openly acknowledged. Conclusions are balanced and supported by the results, without overstating clinical applicability.
Figures and tables are generally well presented and readable. Minor improvements could include harmonizing figure captions stylistically and ensuring consistent terminology (e.g., Ki-67 vs Ki67) throughout.
The manuscript is understandable and scientifically sound, but minor grammatical errors, typographical issues, and occasional awkward phrasing are present. A light professional language editing would improve readability and fluency, particularly in the Discussion section.

Author Response

This manuscript addresses an important and relatively underexplored topic, namely the application of CT-based radiomics and machine-learning approaches for non-invasive Ki-67 stratification in small bowel neuroendocrine tumors (SB-NETs), using a clinically meaningful low-threshold cut-off. Overall, the study is well conceived, methodologically sound, and clearly positioned within the current radiomics literature. The results are promising and potentially clinically relevant.

The introduction provides a solid and coherent background on SB-NET epidemiology, prognostic relevance of Ki-67, and the rationale for radiomics. The gap in the literature is clearly identified, particularly regarding the limited focus on SB-NETs and the use of tailored Ki-67 cut-offs. Relevant and up-to-date references are included. Minor stylistic refinements could further improve clarity and conciseness, but no major conceptual additions are required.

The retrospective, multicenter design is appropriate for an exploratory radiomics study in a rare tumor entity. The lesion-based analysis, although potentially introducing within-patient correlation, is clearly explained and justified given the limited sample size. The choice of a 1% Ki-67 threshold is well motivated and clinically meaningful.

The methodology is described in substantial detail, allowing reproducibility. Image acquisition, segmentation, preprocessing, feature extraction, feature reduction, and machine-learning pipelines are all clearly reported and consistent with IBSI recommendations. The use of cross-validation and SMOTE to address class imbalance is appropriate. One point that could be clarified further is the potential impact of assigning the primary tumor Ki-67 value to metastatic lesions, which is acknowledged later as a limitation.
Results are clearly presented and logically structured. Performance metrics are appropriate and comprehensively reported. Tables and figures are generally clear and informative. The emphasis on recall and false-negative rate is clinically relevant and well justified.
The discussion adequately contextualizes the findings within existing literature, particularly in comparison with pancreatic and rectal NET radiomics studies. Strengths and limitations are openly acknowledged. Conclusions are balanced and supported by the results, without overstating clinical applicability.
Figures and tables are generally well presented and readable. Minor improvements could include harmonizing figure captions stylistically and ensuring consistent terminology (e.g., Ki-67 vs Ki67) throughout.
The manuscript is understandable and scientifically sound, but minor grammatical errors, typographical issues, and occasional awkward phrasing are present. A light professional language editing would improve readability and fluency, particularly in the Discussion section.

 

Thank you very much for your valuable advice. I am currently working on refining the stylistic inconsistencies and improving the fluency of the English text. We are all very excited about the possibility of publishing in this journal!

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Authors have answered most of my queries. The revised manuscript may be accepted for publication.