Fig. 6.

Combination of radiomics and EV PD-L1 dynamics for predicting durable response: (A) Characteristic pipeline for radiomic analysis including CT scan image segmentation, feature extraction, and feature and model selection by machine learning. (B) The introduction of the 6-features radiomic signature into the ΔEV PD-L1 predictive model for RECIST improved its performance as observed in the considerable increase of sensitivity and specificity, with an accuracy of 81.5%. (C) On the contrary, the best model for prediction of irRECIST only included the ΔEV PD-L1 with an accuracy of 74.1% (binary logistic regression)