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

To address the challenges of model instability and limited generalizability in radiomics-based lung cancer prognosis, we developed a robust multiparametric tensor radiomics framework that integrates multiple PET-CT fusion strategies to identify reliable machine learning models and improve the reproducibility and translational potential of imaging-based prognostic modeling. We analyzed 693 lung cancer patients, including 581 for multi-center training and 112 for external testing. PySERA was used to extract 497 IBSI-compliant and moment-invariant radiomics features from PET, CT, and 10 fused PET/CT images, yielding 12 feature flavours per descriptor. Robust features with an intraclass correlation coefficient of at least 0.60 were retained together with clinical variables. Missing data were imputed, ComBat harmonization was applied, and within the tensor radiomics framework, feature flavours were selected by mutual information or combined using an autoencoder. We trained 23 classifiers with 17 feature selection algorithms across individual selected, combined tensor, and concatenated features, as well as clinical features alone. A five-fold semi-supervised strategy using 257 labeled and 324 pseudo-labeled cases was applied for 2-year binary event-free survival prediction, and performance was ranked using a composite performance-stability aware score based on cross-validation metrics. The combined tensor feature model achieved the highest cross-validation stability using ElasticNet feature selection and Extra Trees classifier (composite score 1.947). Performance included balanced accuracy = 0.799 ± 0.007 (external, 0.646); precision = 0.803 ± 0.006 (0.720); recall = 0.800 ± 0.003 (0.714); F1 = 0.801 ± 0.003 (0.717); and ROC-AUC = 0.855 ± 0.024 (0.696). Multiparametric model selection within a tensor radiomics paradigm may enable identification of reliable and generalizable ML models for lung cancer survival outcome prediction.