Censor-aware semi-supervised survival time prediction in lung cancer using clinical and radiomics features

Aim: Lung cancer remains a major global health challenge, and this study presents a censor-aware semi-supervised learning framework (SSL) that integrates clinical and imaging data to improve prognostic modeling and address biases in handling censored data. Methods: We analyzed clinical, positron emission tomography (PET), and computed tomography (CT) data from 199 lung cancer patients from public and local databases, focusing on overall survival time as the primary outcome. Handcrafted (HRF) and Deep Radiomics features were extracted after preprocessing using Visualized & Standardized Environment for Radiomics Analysis (ViSERA) software and were combined with clinical features. Features were reduced using Pearson’s correlation coefficient regression (RR) and the F -test for regression (FR), followed by supervised learning (SL) and SSL. In SSL, censored data were pseudo-labeled using the Weibull accelerated failure time (AFT) model to enrich the training data. Seven regressors and three hazard ratio survival analyses (HRSAs) were optimized using five-fold cross-validation, grid search, and holdout test bootstrapping. Results: For PET-HRFs, the SSL approach reduced the mean absolute error by 14.81%, achieving 1.04 years with FR + AdaBoost Regression (ABR) compared to 1.20 years with SL. For clinical features, SSL with RR + ABR reached a mean absolute error of 1.04 years, outperforming SL (1.09 years) with a 4.9% improvement. In HRSA, CT_HRF combined with principal component analysis (PCA) + Component-Wise Gradient Boosting Survival Analysis yielded an external C-index of 0.65 ± 0.02, effectively distinguishing high- and low-risk groups. Conclusions: The SSL strategy applied to HRFs from PET imaging significantly enhanced survival prediction compared to SL and uncovered complementary biological information that may remain hidden when only limited labeled data are used.