TumorTwin: A python framework for patient-specific digital twins in oncology

TL;DR

TumorTwin addresses the need for a portable, end-to-end digital twin framework in oncology by delivering an open-source Python package that unifies data processing, tumor growth modeling, and decision-support workflows. It introduces a modular architecture with a central PatientData object and interchangeable Model, Solver, and Optimizer components, enabled by GPU-accelerated forward and gradient computations through PyTorch. The framework implements a 3D reaction-diffusion tumor growth model with chemo- and radiotherapy effects, supports gradient-based calibration, and demonstrates feasibility with in silico HGG data (and TNBC in the appendix). This open, extensible platform enables rapid prototyping, cross-site experimentation, and image-guided predictions to inform treatment decisions, with future plans for uncertainty quantification and integration of machine learning components to enhance actionability.

Abstract

Background: Advances in the theory and methods of computational oncology have enabled accurate characterization and prediction of tumor growth and treatment response on a patient-specific basis. This capability can be integrated into a digital twin framework in which bi-directional data-flow between the physical tumor and the digital tumor facilitate dynamic model re-calibration, uncertainty quantification, and clinical decision-support via recommendation of optimal therapeutic interventions. However, many digital twin frameworks rely on bespoke implementations tailored to each disease site, modeling choice, and algorithmic implementation. Findings: We present TumorTwin, a modular software framework for initializing, updating, and leveraging patient-specific cancer tumor digital twins. TumorTwin is publicly available as a Python package, with associated documentation, datasets, and tutorials. Novel contributions include the development of a patient-data structure adaptable to different disease sites, a modular architecture to enable the composition of different data, model, solver, and optimization objects, and CPU- or GPU-parallelized implementations of forward model solves and gradient computations. We demonstrate the functionality of TumorTwin via an in silico dataset of high-grade glioma growth and response to radiation therapy. Conclusions: The TumorTwin framework enables rapid prototyping and testing of image-guided oncology digital twins. This allows researchers to systematically investigate different models, algorithms, disease sites, or treatment decisions while leveraging robust numerical and computational infrastructure.

Figures (6)

• Figure 1: TumorTwin workflow key components.TumorTwin takes as inputs the patient configuration file, MRI data, and an initial guess of model parameters. The patient configuration file and MRI data are used to construct a PatientData object which serves as the central data object throughout TumorTwin. The PatientData object is used to initialize a patient-specific TumorGrowthModel, which can be combined with a ForwardSolver to simulate tumor growth over time. If longitudinal data is available within the PatientData object, an Optimizer module can be used to calibrate model parameters by minimizing the error between model predictions and patient-specific measurements.
• Figure 2: Patient data summary graphics. Example output of the patient data summary function applied to the in silico HGG dataset. This graphic shows treatment and imaging schedule (top panel), anatomical $T_1$-weighted MRI with associated tumor segmentations (middle panel), and the apparent diffusion coefficient ($ADC$) map for the same imaging slices (bottom panel). In practice, this summary can be used to visually confirm longitudinal registration of imaging series, accuracy of tumor segmentations, and treatment details.
• Figure 3: Example tumor growth prediction for the in silico HGG patient. The predicted total tumor cell count over time is shown (top panel), with 2D slices of the model solution at three snapshots in time (bottom panel). Note the invasion, logistic growth, chemotherapy, and radiotherapy effects in the solution.
• Figure 4: Model calibration to patient-specific MRI data. Top: Total tumor cell count (TTC) time series for each iteration of the calibration (black lines; opacity increased with iteration) compared with the observed TTC (purple stars). Middle: Evolution of a central tumor slice across time (left-to-right), and across optimization iterations (top-to-bottom). The bottom row shows the ground-truth data used for calibration and validation. Bottom: Convergence of the loss function over optimization iterations.
• Figure 5: Solver performance profiling for CPU and GPU architectures. Top: Mean and standard deviation of the wall-clock time required to compute a forward model prediction for 180 days. Bottom: Results for a backward adjoint solve for the same period. Two CPU-based and two GPU-based hardware architectures were tested.