DEEP learning in NSCLC demonstrated stronger predictive performance than radiomics models for assessing immunotherapy outcomes in metastatic non-small cell lung cancer, according to findings from an analysis of randomised clinical trials involving programmed cell death protein 1 inhibitors with or without stereotactic ablative body radiotherapy.
Deep Learning in NSCLC Improves Prediction
Researchers investigated whether quantitative imaging approaches including radiomics and deep learning could predict oncological outcomes in metastatic non-small cell lung cancer. The study analysed data from three randomised trials evaluating programmed cell death protein 1 inhibitor therapy with or without stereotactic ablative body radiotherapy: PEMBRO RT, NIVORAD, and MDACC.
A random forest model developed using radiomics features from the PEMBRO RT dataset achieved an area under the curve of 0.57 for predicting per lesion progressive disease during immunotherapy treatment. In comparison, the deep learning model achieved an area under the curve of 0.92, demonstrating substantially improved predictive performance.
The findings highlight the growing potential of artificial intelligence-based imaging analysis to identify treatment response patterns in patients receiving immunotherapy for metastatic non-small cell lung cancer.
Radiomics Models Show Moderate Survival Prediction
Investigators also evaluated radiomics based survival models for overall survival and progression free survival. The random forest survival model achieved a concordance index of 0.63 for overall survival and 0.59 for progression free survival.
Predictive performance improved when clinical variables including programmed death ligand 1 status and treatment arm were incorporated into the model. Following the addition of these clinical features, concordance indices increased to 0.67 for overall survival and 0.65 for progression free survival.
These results suggest that combining imaging derived biomarkers with established clinical parameters may enhance prognostic modelling in metastatic non-small cell lung cancer.
External Validation Remains a Challenge
Validation analyses using the NIVORAD and MDACC datasets demonstrated reduced area under the curve values compared with the original development cohort. The decline in predictive performance across external datasets indicates ongoing challenges with model generalisability.
Although deep learning in NSCLC showed excellent predictive value for identifying per lesion progressive disease in patients receiving immunotherapy, the investigators emphasised that further research is needed before clinical implementation can be achieved.
The authors concluded that improving model generalisability will be essential for successful clinical translation of quantitative imaging tools in metastatic non small cell lung cancer.
Reference
Kothari G et al. Deep learning and radiomics models in patients with advanced non-small cell lung cancer treated with immunotherapy combined with stereotactic radiotherapy. Scientific Reports. 2026; DOI:10.1038/s41598-026-53520-5.
Featured image: Pakorn on Adobe Stock.
