Search papers, labs, and topics across Lattice.
This study introduces an AI-augmented adaptive digital twin framework designed for predicting brain tumor evolution and optimizing treatment schedules. By integrating an interpretable reaction-diffusion model with a 3D residual learning module and model predictive control, the framework significantly enhances prediction accuracy and treatment efficacy. Experiments show that the hybrid model reduces prediction errors by 84.3% and improves treatment outcomes, achieving a 22.4% reduction in final tumor burden compared to fixed schedules, thereby establishing a promising foundation for real-world applications in adaptive treatment planning.
Hybrid modeling techniques can drastically improve the accuracy of brain tumor predictions, reducing errors by over 84% and optimizing treatment schedules for better patient outcomes.
Brain tumor progression exhibits spatially heterogeneous growth, patient-specific treatment response, and complex interactions with surrounding anatomy, making accurate long-term prediction challenging. We propose an AI-augmented adaptive digital twin (DT) framework for brain tumor evolution prediction and treatment scheduling. The framework integrates an interpretable reaction--diffusion (RD) model, a 3D residual learning module for model-form correction, patient-specific DT updating during recursive rollout, and model predictive control (MPC) for constrained chemotherapy and radiotherapy scheduling. Experiments on 387 synthetic tumor trajectories with 120-step evolution show that the baseline RD model captures tumor location and overall temporal behavior but underestimates heterogeneous tumor burden during long-horizon prediction. Hybrid RD--residual modeling reduces masked voxel-wise mean squared error by 84.3% and increases Dice overlap by 43.5% relative to the RD baseline under dense simulated observations. Online DT updating further reduces mean squared error by 45.9% and improves Dice overlap by 9.6% compared with the non-updated hybrid model. In MPC-based scheduling simulations, the updated DT controller reduces final tumor burden by 22.4% relative to a fixed treatment schedule under the terminal-burden objective. Together, these results demonstrate a unified framework for patient-specific initialization, mechanistic modeling, adaptive learning, and constrained treatment optimization. Although validated using patient-data-informed synthetic trajectories rather than clinical longitudinal data, the proposed framework establishes a foundation for future translation to real-world adaptive treatment planning.
