Deep learning and radiomics models in patients with advanced non-small cell lung cancer treated with immunotherapy combined with stereotactic radiotherapy.

Quantitative imaging is an emerging field that may allow prediction of oncological outcomes. We investigate whether radiomics and deep learning can predict outcomes in metastatic non-small cell lung cancer utilizing randomized trials of PD-1 inhibitors + /- stereotactic ablative body radiotherapy: PEMBRO-RT(NCT02492568), NIVORAD(ACTRN12616000352404) and MDACC(NCT02444741). A random forest model developed on PEMBRO-RT using radiomics features had an AUC of 0.57 for prediction of per-lesion progressive disease on immunotherapy compared to an AUC of 0.92 for a deep learning model. A random forest survival model using radiomics features for overall survival (progression free survival) had a concordance index of 0.63(0.59) and improved to 0.67(0.65) by adding clinical features, including PD-L1 and treatment arm. Validation on NIVORAD and MDACC revealed reduced AUCs. Overall, a deep learning compared to a radiomics model demonstrated excellent predictive value for per-lesion progressive disease for patients on immunotherapy. Models had reduced performance on external validation. Research improving generalizability is required for clinical translation.