Multi-Center Semi-Supervised Prediction of Lung Cancer Survival Outcomes Via Pseudolabeling of CT Images

CT imaging is central to lung cancer management, offering detailed tumor visualization and quantitative data for AI-driven personalized care. In this work, we used radiomic features (RFs) and machine learning (ML) to predict overall survival (OS) in non-small cell lung cancer (NSCLC). A total of$1,218 \text{RFs}$were extracted from lung cancer CT images of 977 patients across 14 institutions using LoG and Wavelet filters with varying parameters. Lesions were segmented by three physicians and validated by an expert, with feature extraction standardized using PyRadiomics. Both supervised learning (SL) and semi-supervised learning (SSL) approaches were applied. In SL, 56 dimensionality reduction algorithms-including 27 feature selection (FSA) and 29 attribute extraction (AEA) algorithms-were paired with 29 traditional and advanced classifiers. These combinations were evaluated using 5 -fold cross-validation on the labeled data from one large center (of 417 in NSCLC-Radiomics), with 2 smaller centers (of 34 samples in NSCLC-Radiogenomics and \# of 48 samples in Lung CT-Diagnosis) used for external testing. In SSL, missing outcomes in 11 centers (of 478 samples) were predicted using logistic regression and incorporated into the training set per fold, while unlabeled cases were excluded from validation and external testing. SSL significantly outperformed SL in terms of accuracy, F1-score, and AUC, achieving improvements of up to 17% in OS prediction (paired t-test,$\mathbf{P}<0.01$). The best SSL model-using feature importance and XGBoost-achieved an average cross-validation accuracy of$\mathbf{0. 9 0} \boldsymbol{\pm} \mathbf{0. 0 1}$and an external test accuracy of$0.88 \pm 0.00$. In comparison, the top SL model, recursive feature elimination (RFE) and LightGBM, reached$0.78 \pm 0.02$in cross-validation and$0.87 \pm 0.01$on the external test. These findings highlight SSL's strength in improving NSCLC survival prediction, especially with limited labels, and show the value of multi-center CT RF integration.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.