Multi-Modal PET/CT Fusion for Automated PD-L1 Status Prediction in Lung Cancer

Programmed cell death-ligand 1 (PD-L1) status determination involves surgical or biopsied tumor specimens, which must be collected through invasive procedures that carry a risk of morbidity. Therefore, automating the evaluation of PD-L1 expression from medical images is of paramount importance for patients with lung cancer. This study investigates a variety of deep learning architectures as well as the fusion of non-invasive computed tomography (CT) and positron emission tomography (PET) scans to predict the PD-L1 status. ResNet, DenseNet and EfficientNet models are compared and optimized using different combinations: CT only, PET only and CT/PET fusion. Models are evaluated using the areas under the receiver operating characteristic curves (AUCs) based from their 95% confidence intervals. Results reveal that PET/CT fusion outperforms their independent counterparts whatever the architecture. Further, early fusion demonstrates encouraging results for classifying PD-L1. Such noninvasive, high-throughput deep learning approach can be used as a powerful alternative to broad diagnostic workups, with applications for immunotherapy eligibility assessment.