External Validation of Robust Radiomic Signature to Predict 2-Year Overall Survival in Non-Small-Cell Lung Cancer

Area of Science:

• Radiology and Imaging
• Oncology
• Data Science in Medicine

Background:

• Lung cancer is a leading cause of mortality worldwide.
• Radiomics offers potential for predicting clinical outcomes in lung cancer.
• Feature robustness is a critical challenge for clinical radiomics implementation.

Purpose of the Study:

• To develop and validate robust radiomic signatures for predicting 2-year overall survival in non-small cell lung cancer (NSCLC).
• To address the challenge of radiomic feature robustness in prediction model development.

Main Methods:

• Retrospective study with 300 NSCLC patients (200 internal, 100 external validation from TCIA).
• Radiomic features extracted from CT images; feature selection via hierarchical clustering, Chi-squared test, and RFE.
• Six prediction models (Random Forest, Gradient Boosting, Support Vector) developed and validated using cross-validation, internal, and external validation.

Main Results:

• Top 10 robust radiomic features were selected using a multistep approach.
• Random forest models achieved the highest accuracy: 0.81 (original data) and 0.77 (balanced data) on internal validation.
• External validation showed an accuracy of 0.68 for the random forest models, indicating promising generalizability.

Conclusions:

• Robust radiomic features demonstrate significant potential for predicting 2-year overall survival in NSCLC patients.
• The developed radiomic signatures, particularly those from random forest models, offer a promising tool for lung cancer prognosis.
• Further validation is warranted to fully integrate these radiomic approaches into clinical practice for lung cancer management.