Toward Actionable Digital Twins for Radiation-Based Imaging and Therapy: Mathematical Formulation, Modular Workflow, and an OpenKBP-Based Dose-Surrogate Prototype

A team of researchers has introduced a new, modular framework for building 'actionable digital twins' in medical radiation therapy, moving beyond static simulations to systems that can assimilate patient data and support clinical decisions. The framework, detailed in a new arXiv paper, formalizes a workflow coupling PatientData, Model, Solver, Calibration, and Decision modules. It focuses on latent-state updating, uncertainty propagation, and chance-constrained action selection, aiming to make digital twins practical tools for personalized, adaptive therapy.

The researchers instantiated their framework with a concrete, open-source prototype using the OpenKBP benchmark dataset. This implementation features a GPU-ready, 19.2-million-parameter 3D U-Net built in PyTorch/MONAI. The model is trained with a specialized masked loss and is equipped with Monte Carlo dropout to provide voxel-wise epistemic uncertainty estimates—a crucial feature for clinical trust. In tests, the system achieved a mean dose score of 2.65 Gy on a 100-patient test set. Most impressively, it executed a full three-fraction treatment simulation loop, including model recalibration and spatial optimization, in just 10.3 seconds, with a mean inference time of 0.58 seconds per patient.

This work establishes a reproducible test bed for dose prediction and uncertainty-aware therapy evaluation. The speed and modularity of the framework pave the way for integrating real-world data streams, like delivered-dose logs and repeat imaging, to create digital twins that continuously learn and adapt to individual patient responses over the course of treatment.