TumorGrowth.jl is a new package providing ODE-based models for predicting the changes in volume of a lesion. In addition to classical models, like Bertalanffy and Gompertz, it includes, as experimental, some neural ODE models. It has tools to calibrate the model parameters, given patient measurements, and a tool to compare multiple models on a holdout set.
Calibrating other ODE’s
A lot of the code-base (including the neural ODEs) is generic for one-dimensional, temporal ODE’s and so could be adapted to calibrate and compare such models in other scientific contexts, for those comfortable with Julia programming. Optimization is either using ordinary gradient descent, or Newton-Gauss methods (Marquardt-Levenberg or dog leg). In the case of gradient descent, flexible iteration “controls” (such as early stopping criteria) are provided by IterationControl.jl. We incorporate SciML tools to solve the ODEs, hooking into Pontryagin’s adjoint method for efficiently differentiating solutions with respect to parameters and initial conditions.