Conformalized Semi-Supervised Learning for Medical Imaging

To investigate whether conformal prediction (CP)-based uncertainty quantification (UQ) can improve semi-supervised learning (SSL) frameworks. Five machine learning (ML) frameworks were developed and evaluated: supervised learning (SL), standard SSL, iterative SSL, conformalized SSL (CSSL), and iterative CSSL. Logistic regression (LR) models were trained using radiomics features extracted from MRI and CT datasets across three classification tasks: isocitrate dehydrogenase (IDH) mutation prediction in gliomas, O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation prediction in gliomas, and histopathological classification in non-small cell lung cancer (NSCLC). Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC) with 95% confidence intervals, and statistical comparisons were performed using DeLong’s test. Across all three tasks, uncertainty-aware SSL frameworks outperformed conventional SL and standard SSL approaches. For IDH prediction, CSSL and iterative CSSL demonstrated statistically significant improvements compared with SL and SSL ( p < 0.05). For MGMT prediction, iterative SSL, CSSL, and iterative CSSL significantly outperformed SL ( p < 0.05). Similarly, in NSCLC histology classification, iterative SSL and CSSL-based methods significantly improved performance compared with SL ( p < 0.05). Integrating CP into SSL frameworks improves pseudo-label reliability by filtering uncertain predictions and reducing error propagation during self-training. The proposed uncertainty-aware SSL approaches demonstrated improved performance across multiple radiomics-based classification tasks, highlighting their potential utility in medical imaging applications where labeled data are limited but unlabeled data are abundant.