Feature Tokenizer-Transformers with Self-Training for The Prediction of PD-L1 Expression of Non-Small Cell Lung Cancer from CT

Patients with advanced non-small cell lung cancer (NSCLC) have notoriously unfavorable clinical outcomes. Immune checkpoint inhibitors (ICIs) have recently shown promising results in the treatment of this malignancy by targeting immune checkpoint proteins such as the programmed death ligand-1 (PD-L1). Nevertheless, anti-PD-1 inhibitors are effective in less than half of the NSCLC patients, and are subsequently administered alone as the first-line of treatment only when the PD-L1 expression is above the 50% threshold. The objective of this work is to non-invasively classify PD-L1 expression levels within a lung nodule using pre-treatment CT images, largely available in NSCLC diagnosis. To this end, we propose a pipeline based on the Feature Tokenizer-Transformer (FT-Transformer) and self-training by leveraging both labeled and unlabeled data. Experimental results showed that the self-trained FT-Transformer exhibited a good predictive performance (accuracy = 0.73, AUC = 0.75) and outperformed traditional machine learning models. Finally, we demonstrate that the predicted Rad-PDL1 score is associated with the patients’ clinical outcomes, namely the progression free survival (p = 0.02).