Semi-Supervised Learning for Improved Radiomics-Based Outcome Prediction from Lymphoma [${ }^{18}\mathrm{F}]\text{FDG}$ PET Images

PET imaging is integral to lymphoma management, offering valuable functional insights and quantitative biomarkers towards personalized treatment planning. In this study, we utilized radiomics features (RF) and machine learning (ML) to predict Relapse-Free Survival (RFS) in lymphoma patients. A total of 484 RFs were extracted from PET scans of 776 patients across two lymphoma datasets, including 631 samples from the 2024-SNMMI-Challenge (296 with and 335 without RFS labels) and 145 from AutoPET (unlabeled). RF extraction was performed using the RaCaT software. The labeled data were partitioned, with 80% used for 5-fold cross-validation and the remaining 20% reserved for final testing. We explored both supervised learning (SL) and semisupervised learning (SSL) strategies. The SL framework evaluated 56 dimensionality reduction methods (set to reduce feature size to 10), comprising 27 feature selection algorithms (FSA) and 29 attribute extraction algorithms (AEA), in combination with 29 classical and advanced classifiers. In the SSL framework, unlabeled samples from two additional datasets (480 patients) were pseudo-labeled using logistic regression and integrated into the training set within each fold. These pseudo-labeled cases were excluded from both validation and final testing. SSL consistently outperformed SL in terms of accuracy, F1-score, and AUC, with improvements reaching up to 14% (paired ttest, p-value <0.01). The best-performing SSL model, incorporating the ENet feature selector and MLP classifier, achieved an average cross-validation accuracy of 0.925 ± 0.003 and an external test accuracy of 0.811 ± 0.000. In comparison, the top SL model, with the same ML algorithms, achieved 0.809 ± 0.045 in cross-validation and 0.787 ± 0.018 in final testing. These findings emphasize the superiority of SSL in scenarios with limited labeled data and demonstrate the value of integrating multicenter PET-based RFs for robust survival prediction in lymphoma.

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.