Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript presents a radiomics-based AI classification model to differentiate between spinal tuberculosis (TB) and spinal metastasis (METAS) using MRI images. However, some problems should be addressed as follows:

1. The study is based on only 50 patients (25 TB, 25 METAS). Although the authors use augmentation and cross-validation to compensate, the small original sample size remains a fundamental limitation. The external validation set (n=10) is also very small.

2. Segmentation was performed manually by radiologists and orthopedists with 4–5 years of experience, using consensus. However, no inter-rater reliability metric (e.g., Dice coefficient or Intraclass Correlation Coefficient) is reported. This is critical because radiomics features are highly sensitive to segmentation variability.

3. The top 10 features were selected using XGBoost-SHAP on the full dataset (n=50). This is a pre-selection bias because feature selection was performed before cross-validation, potentially leaking information from the test set.

4. Figure 9 shows that SFA and SMOTE yield nearly identical results, and performance saturates at X3. However, the authors do not explain why SMOTE and SFA perform similarly or why X3 is optimal.

5. Table 4 shows individual doctor performance, but no statistical comparison (e.g., paired t-test, McNemar's test) is performed between the AI model and human experts. The claim that AI "outperforms" humans is not statistically supported. Some related references should be addressed as follows: Dong S, Wang P, Abbas K. A survey on deep learning and its applications[J]. Computer Science Review, 2021, 40: 100379.

6. The external validation set (10 cases, 5 TB + 5 METAS) comes from two independent hospitals. However, the authors do not provide:

• The acquisition parameters for these external images,

• Whether the same segmentation protocol was applied,

• The performance breakdown per class (confusion matrix for external set).

Author Response

Please see the attached file.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1.The sample design exhibits temporal bias, necessitating supplementary discussion and optimization of enrollment balance. TB cases include all eligible patients from 2019–2021, while METAS only selected the latest cases from 2021, resulting in inconsistent data collection time windows between the two groups. Additionally, TB cases primarily involve MRI scans from multiple external hospitals and vendors, whereas METAS data almost exclusively comes from Siemens equipment at our institution, leading to imbalanced equipment and source distribution. This introduces stratified confounding bias, which affects the model's real-world generalization conclusions.

2.The sample size of the external validation set is too small, and the generalization conclusion needs to be expressed cautiously. The external validation only had 10 cases (5TB+5METAS), with an extremely low sample size and a wide statistical confidence interval of 80% accuracy, which cannot effectively prove the model's cross center stability; There were only 2 false positives (diagnosed as TB by METAS), and the imaging characteristics of misdiagnosed cases were not analyzed. There was a lack of review of failed cases.

3.The feature augmentation method is only based on eigenvalue perturbation and lacks image level augmentation control. It only performs Gaussian perturbation augmentation (SFA) on the extracted numerical features and does not use the image spatial augmentation commonly used in medical imaging (rotation, translation, elastic deformation, intensity perturbation); Unable to distinguish the advantages and disadvantages of "feature augmentation" and "image augmentation", and incomplete demonstration of small sample optimization.

4.Lack of horizontal performance comparison with existing spinal identification AI models. The article only compares with human physicians and does not search for similar AI identification literature on spinal tuberculosis/metastasis imaging in recent years. There is no horizontal index comparison, which cannot highlight the improvement of this model compared to previous work.

5.The risk assessment dimension of model overfitting is insufficient, relying only on cross validation and feature dimensionality reduction to alleviate overfitting, without drawing training/validation loss curves and accuracy convergence curves; The overfitting inflection point of the training epochs has not been analyzed, making it difficult to visually demonstrate the effectiveness of the early stopping strategy.

Author Response

Please see the attached file.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

In this work, authors have presented an AI/ML pipeline to classify  or distinguish TB and METAS  using feed forward neural net trained on radiomics features extracted from MRI scans. The work is interesting and addresses a real world problem. However, there are a few points that would make the work stronger.
1. Authors have used a feed forward neural network (FFNN). Could authors discuss their choice as compared to a 3D CNN  (like a 3D resnet or UNET)? 
2. Related to ML architecture- did authors benchmark the FFNN against other standard classifiers like SVMs and Random forest to evaluate if deep networks offer an advantage on the feature size they are using?
3. Could the authors please provide distribution plots or violin plots for top 10 features comparing the TB and METAS cluster cohorts? Was there a statistical testing performed to evaluate whether the extracted features have statistically significant discriminative power?
4.Could the authors also provide the top ranking features according to their feature importance?
5. Regarding feature selection- did authors perform any collinearity filtering or analysis to determine and eliminate any redundant features? if so, please provide that information in the manuscript.

6. Training a multi-layer neural net on just 50 cases presents a risk of overfitting. How have authors addressed this issue?

7. Although SMOTE and SFA has been applied, it is still based on those 50 cases. It would not relaistically cover true clinical variance. Could authors provide learning curves (loss/accuracy over epochs) to show baseline stability?

8. Have authors considered testing the model against background/control group or non-TB/non-METAS spinal pathologies to evaluate false-positive rates?

Author Response

Please see the attached file.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The paper has addressed all my questions.

Author Response

Thank you very much for your valuable comments and suggestions. Your feedback has helped us improve the quality of our manuscript.

Reviewer 2 Report

Comments and Suggestions for Authors

The author has provided detailed responses and corresponding revisions to the reviewers' comments, and it is recommended to accept this version of the manuscript

Author Response

Thank you very much for your valuable comments and suggestions. Your feedback has helped us improve the quality of our manuscript.