An <scp>MRI</scp>‐based machine learning radiomics can predict short‐term response to neoadjuvant chemotherapy in patients with cervical squamous cell carcinoma: A multicenter study

Zhonghong Xin, Wanying Yan, Yibo Feng, Yunzhi Li, Yaping Zhang, Dawei Wang, Weidao Chen, Jianhong Peng, Cheng Guo, Zixian Chen, Xiaohui Wang, Jun Zhu, Junqiang LeiCancer 2023 · DOI

TL;DR

This study developed and validated an MRI-based machine learning radiomics model to predict short-term response to neoadjuvant chemotherapy (NACT) in cervical squamous cell carcinoma (CSCC) patients. The model, particularly using Support Vector Machines (SVM), demonstrated promising accuracy and robustness across a multicenter dataset, offering a non-invasive tool to guide personalized treatment decisions.

What Problem Does This Paper Address?

Cervical squamous cell carcinoma (CSCC) is a significant cause of cancer-related deaths in women. Neoadjuvant chemotherapy (NACT) is a crucial treatment component, aiming to shrink tumors and improve surgical outcomes. However, a substantial portion of patients (30-50%) do not respond well to NACT due to individual differences in sensitivity. Current methods for assessing NACT response, such as histopathology, are invasive and can delay treatment. While imaging techniques like MRI are non-invasive, their interpretation can be subjective, and conventional methods struggle to accurately predict treatment response. This gap highlights the need for an objective, non-invasive, and accurate method to predict NACT efficacy, enabling personalized treatment strategies and avoiding ineffective therapies.

Methods

The researchers employed a machine learning radiomics approach. Key steps included:

  • Image Acquisition: Standardized pelvic MRI scans (axial T2WI sequences) were performed before and after NACT.
  • Region of Interest (ROI) Segmentation: A radiologist manually delineated the tumor region on axial T2WI images using a specialized platform.
  • Radiomic Feature Extraction: A comprehensive set of radiomic features (over 1400 initially) were extracted from the segmented ROIs, encompassing categories like first-order, gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), gray level dependence matrix (GLDM), neighboring gray tone difference matrix (NGTDM), and shape features.
  • Feature Selection: To reduce dimensionality and improve model performance, Recursive Feature Elimination (RFE) was used to select the most informative 10 features.
  • Model Development: Three machine learning algorithms – Logistic Regression (LR), Random Forest (RF), and Support Vector Machines (SVM) – were trained to predict NACT susceptibility.
  • Model Evaluation: The performance of each model was assessed using Receiver Operating Characteristic (ROC) curve analysis (Area Under the Curve - AUC), accuracy, sensitivity, and precision.

Datasets

The study utilized a multicenter dataset comprising 234 patients with CSCC. These patients were divided into:

  • Training Set: 180 patients
  • Testing Set: 20 patients
  • External Validation Set: 34 patients

The data was collected from two hospitals between January 2019 and May 2022.

Key Insights and Ideas

  • MRI T2WI as a Predictive Tool: The study successfully demonstrated that axial T2WI MRI sequences, often overlooked for detailed radiomic analysis, contain valuable information for predicting NACT response in CSCC.
  • Radiomics Captures Heterogeneity: Radiomics, by extracting quantitative features from images, offers a way to capture tumor heterogeneity across the entire volume, which is crucial for understanding treatment response, unlike biopsy which samples only a small subset.
  • SVM's Superior Performance: Among the tested machine learning algorithms, SVM consistently showed the most robust performance in predicting NACT response, both in the testing and external validation sets.
  • Multicenter Validation is Crucial: The inclusion of an external validation set from a different center strengthens the generalizability and clinical applicability of the developed model, addressing a key limitation in previous radiomics studies.

Key Results

The SVM model achieved the highest predictive performance:

  • Testing Set:
    • AUC: 0.88 (95% CI: 0.687–1.0)
    • Accuracy: 0.90 (95% CI: 0.75–1.0)
    • Sensitivity: 0.933 (95% CI: 0.778–1.0)
  • External Validation Set:
    • AUC: 0.764 (95% CI: 0.558–0.947)
    • Accuracy: 0.853 (95% CI: 0.735–0.971)
    • Sensitivity: 0.962 (95% CI: 0.875–1.0)

While LR and RF also showed high accuracy in the testing set (1.0), SVM's balanced performance across metrics and its strong showing in the external validation set highlight its potential.

Limitations and Future Work

The authors acknowledge several limitations:

  • Retrospective Data and Inherent Bias: The retrospective nature of the study and data from different centers, despite efforts, may introduce biases.
  • Single Imaging Sequence: The study primarily used axial T2WI images. Future work could explore other MRI sequences (e.g., DWI) or combine multiple sequences for enhanced predictive power.
  • Exclusion of Clinical Factors: The model did not incorporate clinical risk factors, which have been shown to influence NACT prediction.

Future directions suggested include:

  • Prospective Multicenter Studies: Conducting prospective studies with larger, more diverse patient cohorts from various regions to further validate and generalize the model.
  • Multi-sequence MRI Fusion: Integrating radiomic features from different MRI sequences (T1WI, DWI, etc.) to create a more comprehensive predictive model.
  • Integration of Clinical Data: Developing a joint prediction model that combines imaging and clinical data to improve predictive accuracy and provide more personalized treatment insights.

Summary updated

Abstract

Machine learning radiomics models based on MRI images have achieved satisfactory performance in predicting the sensitivity of NACT in CSCC patients with high accuracy and robustness, which has great significance for the treatment and personalized medicine of CSCC patients.
BibTeX
@inproceedings{xin2023scp,
  author = {Zhonghong Xin and Wanying Yan and Yibo Feng and Yunzhi Li and Yaping Zhang and Dawei Wang and Weidao Chen and Jianhong Peng and Cheng Guo and Zixian Chen and Xiaohui Wang and Jun Zhu and Junqiang Lei},
  title = {{An <scp>MRI</scp>‐based machine learning radiomics can predict short‐term response to neoadjuvant chemotherapy in patients with cervical squamous cell carcinoma: A multicenter study}},
  booktitle = {Cancer},
  year = {2023},
  doi = {10.1002/cam4.6525},
}
Figure