Indirect supervised fine-tuning of a tumor model parameter estimator neural network

Abstract

Personalized therapy based on mathematical foundations is a promising method for treating various types of cancer. By identifying the parameters of mathematical equations, we could gain more information about the patients and the tumor. In previous works, a vast number of training data of virtual scenarios has been generated, then used to train a neural network to predict parameters. Besides the fact that in silico experiments have been used, and the actual parameters can differ from them, the algorithm still can be utilized for initial estimation. The main objective of this work is to find the parameters of living mice, by taking advantage of the learning capability of neural networks. As a result, the implementation encompasses two main stages. First, we created another supervised neural network, that is able to solve the applied differential equations faster with fewer algebraic steps, than the traditionally used ODE solvers. Then, we find a better-fitting parameter set for the real measurement, while we retrain the original network with these parameters and the associated error, without forgetting the already learned weights from in silico experiments. The results indicate that the implemented model can be used in further research as an unconstrained optimization technique for parameter fitting.