Semi-Supervised and Unsupervised Deep Learning Combination for Automated PDL-1 Status Prediction in Lung Cancer with Multi-modal PET/CT Fusion

Improving the amount of labeled information and the quality of labels is essential for any deep neural networks (DNNs) based classification task using PET/CT images. In addition, high-quality annotation depends on the experience of imaging specialists, introducing a certain variability in the quality level of achieved annotations. In this context, our aim was a combination of semi-supervised and unsupervised deep neural networks using early fusion multi-modal PET/CT images for the prediction of programmed death ligand-1 (PDL-1) status in non-small cell lung cancer (NSCLC). Models were assessed using areas under the receiver operating characteristic curves (AUCs) considering 95% confidence intervals (CI). Compared with current methods, the proposed methodology enhances the robustness of the model and reduces the influence of outliers. The framework also yielded better performance when compared to current methods for PD-L1 classification, consistently outperforming current methods when using different levels of unlabelled PET/CT images. The results show the effectiveness of combining supervised and unsupervised DNNs for handling data with heterogeneous levels of labelled information.