A multi-task masked autoencoder with GAN-based augmentation for PD-L1 prediction from chest CT images

Targeted and immune-based therapies, such as PD-1/PD-L1 inhibitors, have become standard treatments for advanced non-small cell lung cancer (NSCLC). However, accurately identifying patients who benefit from these therapies remains challenging due to tumor heterogeneity and variability in PD-L1 staining. To address this issue, we propose a non-invasive, model-driven computer-aided diagnosis framework that predicts PD-L1 expression directly from CT images under limited labeled data conditions. This study included 188 NSCLC patients from two university hospitals, of whom 49 had PD-L1 expression ≥ 50% and 139 had < 50%. We introduce a Multi-task Masked Autoencoder (MTMAE) with three key components: (1) a self-supervised masked image modeling strategy to leverage unlabeled data and improve data efficiency, (2) an integrated segmentation task to enhance tumor-focused feature learning, and (3) a Gabor-based generative adversarial network for data augmentation to improve generalization. The proposed model achieved an AUC of 0.735 and an accuracy of 0.724, outperforming traditional supervised pretraining (AUC 0.695) and single-task MAE (AUC 0.712). These results demonstrate that combining self-supervised learning, multi-task learning, and GAN-based augmentation enables a reproducible and standardized model-based prediction of clinically reported PD-L1 status from CT images, providing a non-invasive complementary tool for treatment decision-making.