AI-Powered Lesion-Level Tumor Growth Inhibition Modeling Improves Model Stability and Prognostic Association With PFS

Response Evaluation Criteria in Solid Tumors (RECIST 1.1) is currently the standard for tumor response measurement, but it is time-consuming, labor-intensive, and subject to reader variability. Tumor growth inhibition (TGI) models built on RECIST datasets often consider the time-course of all lesions as one aggregated tumor, as represented by the sum of longest diameters. An AI-based tool was utilized to obtain comprehensive data of both aggregated tumor and individual lesion growth dynamics in terms of their longest diameters and 3D volumes. Standard and AI-derived measurements were used to develop TGI models to evaluate the time-course of aggregate and lesion-level tumor shrinkage for patients receiving lorlatinib in the phase III CROWN trial. Additionally, a parametric time-to-event model was developed to describe the PFS probability. Lesion-level TGI models were connected to the PFS model by investigating the influence of various TGI-derived tumor metrics and parameter estimates on the association with PFS. TGI models developed using the more comprehensive AI-generated lesion measurements demonstrated better precision and were more robust than traditional RECIST-based models when tested under different scenarios using stochastic simulation and estimation (SSE). The mean value of tumor decay rate (K_D) and the mean lesion size at week 8 were found to be significantly associated with PFS from the lesion-level models. In contrast, aggregated TGI models did not identify any significant prognostic factors of PFS. This modeling approach, utilizing AI-powered auto-measurement and incorporating individual lesion dynamics, provides a more complete understanding of tumor growth and can enable early decision-making for lesion-targeting therapies.