CT radiomics-based model for predicting TMB and immunotherapy response in non-small cell lung cancer

Jiexiao Wang, Jialiang Wang, Xiang Huang, Yanfei Zhou, Jian Qi, Xiaojun Sun, Jinfu Nie, Zongtao Hu, Shujie Wang, Bo Hong, Hongzhi WangBMC Medical Imaging 2024 · DOI

TL;DR

This study introduces a novel CT radiomics-based model that accurately predicts Tumor Mutational Burden (TMB) and immunotherapy response in non-small cell lung cancer (NSCLC) patients. By analyzing subtle CT image features, this non-invasive approach offers a cost-effective alternative to expensive genomic sequencing for identifying patients likely to benefit from immunotherapy.

What Problem Does This Paper Address?

Immunotherapy, particularly immune checkpoint inhibitors, has revolutionized cancer treatment, but its effectiveness varies significantly among patients. Identifying which NSCLC patients will respond to these therapies is a major challenge. While Tumor Mutational Burden (TMB) is a promising biomarker for predicting immunotherapy response, its measurement typically requires expensive and time-consuming whole-exome sequencing (WES). This creates a barrier to widespread clinical application, especially for patients with advanced or inoperable disease. The paper aims to bridge this gap by developing a non-invasive, cost-effective method using readily available CT scan data to predict TMB and, consequently, immunotherapy response.

Methods

The researchers employed a multi-stage approach:

  • Radiomics Feature Extraction: They used the 3D-Slicer software to extract over 1000 radiomic features from pre-operative chest CT images of NSCLC patients. These features fall into categories like morphological, first-order histogram, texture, and wavelet-based features.
  • Feature Selection: To manage the high dimensionality of radiomic data and prevent overfitting, they utilized the Least Absolute Shrinkage and Selection Operator (LASSO) regression. This technique identified three key radiomic features: Flatness, Autocorrelation, and Minimum.
  • Model Development: A multivariate logistic regression algorithm was used to build a predictive model for TMB levels based on the selected radiomic features. This model was visualized as a nomogram.
  • Validation: The model's performance was rigorously evaluated using two independent external validation datasets. Receiver Operating Characteristic (ROC) curves and calibration curves were used to assess predictive accuracy and agreement between predicted and observed probabilities. Decision Curve Analysis (DCA) was employed to evaluate the clinical utility and net benefit of the model.

Datasets

  • Training Dataset: 62 NSCLC patients from The Cancer Genome Atlas (TCGA) database, with available chest CT images and TMB data.
  • Validation Dataset I: 18 NSCLC patients from Hefei Cancer Hospital (HFCH), Chinese Academy of Sciences, with available CT images and targeted NGS data for TMB estimation.
  • Validation Dataset II: 47 NSCLC patients from HFCH, with available CT images and immunotherapy response data.

Key Insights and Ideas

  • CT Images Hold Predictive Power: The study demonstrates that subtle, quantitative features within standard CT scans can capture information related to a tumor's biological characteristics, such as TMB.
  • Three Features are Key: Surprisingly, a robust predictive model for TMB could be built using just three radiomic features (Flatness, Autocorrelation, and Minimum), suggesting these features are highly informative and potentially reduce the risk of overfitting compared to models using many features.
  • Non-Invasive Biomarker Potential: The radiomics model offers a non-invasive and cost-effective alternative to genomic sequencing for predicting TMB, making it more accessible for clinical decision-making.
  • TMB as a Proxy for Response: The study reinforces the link between higher TMB and better immunotherapy response, and their model effectively predicts this response.

Key Results

  • TMB Prediction (Training Set): The radiomics model achieved an Area Under the Curve (AUC) of 0.816 (95% CI: 0.7109–0.9203) for predicting TMB levels in the TCGA training dataset.
  • TMB Prediction (Validation Set I): The model demonstrated good performance in the first external validation set with an AUC of 0.775 (95% CI: 0.5528–0.9972).
  • Immunotherapy Response Prediction (Validation Set II): The model showed a satisfactory AUC of 0.762 (95% CI: 0.5669–0.9569) for predicting immunotherapy efficacy.
  • Clinical Utility: Decision Curve Analysis indicated that the radiomics model provides a net benefit for clinical intervention across a range of threshold probabilities, suggesting its practical applicability.

Limitations and Future Work

  • Retrospective and Single-Center Study: The study is retrospective and primarily validated on Chinese patients from a single medical center. Further validation in large-scale, multi-center, multi-ethnic prospective randomized clinical trials is crucial.
  • Manual Segmentation: The manual segmentation of the region of interest is time-consuming and labor-intensive. Development of an automated segmentation algorithm is suggested for future work.
  • Unexplained Biological Mechanisms: The precise biological mechanisms linking radiomic features to TMB remain largely unexplained and require further investigation.
  • Generalizability: While the model shows promise, its generalizability across diverse patient populations and scanner types needs to be confirmed.

Summary updated

BibTeX
@article{wang2024ct,
  author = {Jiexiao Wang and Jialiang Wang and Xiang Huang and Yanfei Zhou and Jian Qi and Xiaojun Sun and Jinfu Nie and Zongtao Hu and Shujie Wang and Bo Hong and Hongzhi Wang},
  title = {{CT radiomics-based model for predicting TMB and immunotherapy response in non-small cell lung cancer}},
  journal = {BMC Medical Imaging},
  year = {2024},
  doi = {10.1186/s12880-024-01221-8},
}
Figure