Parameter identification of a tumour growth model for adaptive radiotherapy

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

Adaptive Radiation Therapy (ART) is a crucial medical technique in the field of Radiation Therapy (RT). It aims to dynamically adapt the treatment plan using systematic measurements. This master thesis investigates the possibility of creating an ART process in which a tumour growth model adjust itself to the real tumour anatomy. A new treatment plan can then be optimised using Reinforcement Learning (RL) as demonstrated by Moreau et al. and Martin et al. . Specifically, we address the inverse parameter identification problem on the combined cellular simulation proposed by Jalalimanesh et al. and O’Neil et al. . It is achieved by training a hybrid Neural Network (NN) architecture combining Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) on the output matrices of the simulation. In practise, we are trying to predict five crucial initial parameters of the simulation: the duration of the cell cycle and the four average consumptions in both oxygen and glucose for both cancer and healthy cells. Our research shows that some of these parameters can be accurately predicted while others, which have less impact on the simulation, are more challenging to estimate. Notably, our most promising result is predicting the duration of the cell cycle with an error margin of within 5%. Furthermore, due to the stochastic nature of the tumour growth model, we analyse the inherent similarity between simulations. This leads us to observe the consequences of the prediction error on the output matrices of the model. This master thesis demonstrates the potential to dynamically adjust the parameters of a cellular model based on its outputs, leading to new prospects in the field of ART.