Fig. 1

Predictive models for durable response: (A) Radiomic feature map sample of different expression of a radiomics feature (TL_F101: Range of Difference Variance) in non-responder and responder lesions from baseline computed tomography scans. (B) The predictive model of dynamics of ΔEV PD-L1 + Radiomics showed a 76.7% area-under-the-curve (AUC) to predict non-responders between patients receiving Pembrolizumab + Docetaxel, revealing an improvement in 15% over the Radiomics model and ~ 14% over the tissue PD-L1. To the contrary, only 56.3% AUC was found for ΔEV PD-L1 + Radiomics in the Docetaxel group, which was pretty similar to the one associated with tissue PD-L1 (binary logistic regression)