Automatic Feature Construction Based on Genetic Programming for Survival Prediction in Lung Cancer Using CT Images

Area of Science:

• Radiomics
• Machine Learning
• Medical Imaging Analysis

Background:

• Radiomics analysis uses machine learning for classification models.
• Irrelevant and redundant features can decrease classification performance.
• Effective feature selection is crucial for accurate medical predictions.

Purpose of the Study:

• To propose Genetic Programming (GP) for automatic construction of reduced, relevant radiomic features.
• To enhance classification performance in Non-Small Cell Lung Cancer (NSCLC) prediction.
• To improve patient stratification based on overall postoperative survival.

Main Methods:

• Application of the Genetic Programming (GP) algorithm to select independent and relevant radiomic features.
• Utilizing pre-operative computed tomography (CT) images from NSCLC patients.
• Developing linear classifiers to predict two-year survival.

Main Results:

• The GP-based model demonstrated superior classification performance ([Formula: see text]) compared to benchmark models ([Formula: see text] and 0.64).
• The proposed method successfully reduced the number of radiomic features while maintaining relevance.
• Improved accuracy in stratifying patients by high and low risk of mortality.

Conclusions:

• Genetic Programming offers a powerful tool for optimizing radiomic feature selection.
• The GP-based approach enhances the predictive accuracy of survival in NSCLC patients.
• This method holds promise for personalized risk stratification in cancer care.