Multiparametric PET/CT Tensor Radiomics for Stability-Aware Machine Learning in Lung Cancer Survival Prediction

Area of Science:

• Medical Imaging
• Radiomics
• Machine Learning
• Oncology

Background:

• Radiomics shows promise in lung cancer prognosis but faces challenges with model instability and limited generalizability.
• Integrating multiple imaging modalities and advanced computational frameworks is crucial for improving prognostic accuracy.

Purpose of the Study:

• To develop and validate a robust multiparametric tensor radiomics framework for lung cancer prognosis.
• To enhance the reproducibility and translational potential of imaging-based prognostic models by addressing instability and generalizability issues.

Main Methods:

• A cohort of 693 lung cancer patients (581 training, 112 external testing) was analyzed.
• Radiomics features were extracted from PET, CT, and fused PET/CT images using the PySERA tool, with 12 feature flavors per descriptor.
• A tensor radiomics framework combined feature flavors, applied feature selection, and trained 23 classifiers for 2-year event-free survival prediction using a semi-supervised strategy.

Main Results:

• The combined tensor feature model, utilizing ElasticNet and Extra Trees classifier, demonstrated the highest cross-validation stability (composite score 1.947).
• The model achieved a balanced accuracy of 0.799 ± 0.007 (external 0.646) and ROC-AUC of 0.855 ± 0.024 (external 0.696).
• The framework successfully integrated multiparametric data and advanced machine learning for reliable prognostic predictions.

Conclusions:

• Multiparametric model selection within a tensor radiomics framework can identify reliable and generalizable machine learning models for lung cancer survival prediction.
• This approach holds potential for improving the clinical utility of radiomics in oncology.
• The study highlights the importance of robust feature selection and model validation for translational research in medical imaging.