Variogram Modeling of Spatially Variant Early Response to Therapy in Advanced Non-Small Cell Lung Cancer.

Abstract

Predicting heterogeneity in treatment response for non-small cell lung cancer (NSCLC) at multiple scales, both between patients and spatially within each patient, can support clinical decisions that personalize oncologic management. In this study, we evaluated different variogram models of voxel-level spatial correlation in tumor response in locally advanced NSCLC and metastatic NSCLC from two different clinical trials. The Stable model achieved the lowest root mean squared error (RMSE) on average (mean: 5.2-5.5%), followed by the Matérn model (mean: 5.8-7.4%), both of which performed better than most other models. In contrast, the Exponential model had the highest RMSE (mean: 9.4-15.6%). These results remained consistent across two different cohorts of NSCLC. Given the robust performance of the Stable model, it may generalize for modeling spatial response in other clinical settings beyond NSCLC and should be further studied.