Identifying early tumour states in a Cahn-Hilliard-reaction-diffusion model

TL;DR

This paper describes the tumour evolution through a diffuse interface model coupling a Cahn-Hilliard-type equation for the tumour phase field to a reaction-diffusion equation for a key nutrient proportion, also accounting for chemotaxis effects and proposes a Tikhonov regularisation approach that approximates the solution through a family of constrained minimisation problems.

Abstract

In this paper, we tackle the problem of reconstructing earlier tumour configurations starting from a single spatial measurement at a later time. We describe the tumour evolution through a diffuse interface model coupling a Cahn-Hilliard-type equation for the tumour phase field to a reaction-diffusion equation for a key nutrient proportion, also accounting for chemotaxis effects. We stress that the ability to reconstruct earlier tumour states is crucial for calibrating the model used to predict the tumour dynamics and also to identify the areas where the tumour initially began to develop. However, backward-in-time inverse problems are well-known to be severely ill-posed, even for linear parabolic equations. Moreover, we also face additional challenges due to the complexity of a non-linear fourth-order parabolic system. Nonetheless, we can establish uniqueness by using logarithmic convexity methods under suitable a priori assumptions. To further address the ill-posedness of the inverse problem, we propose a Tikhonov regularisation approach that approximates the solution through a family of constrained minimisation problems. For such problems, we analytically derive the first-order necessary optimality conditions. Finally, we develop a computationally efficient numerical approximation of the optimisation problems by employing standard $C^0$-conforming first-order finite elements. We conduct numerical experiments on several pertinent test cases and observe that the proposed algorithm consistently meets expectations, delivering accurate reconstructions of the original ground truth.

Figures (10)

• Figure 1: Plots of the initial condition and the final condition at $t=30$ days, for both the variables $\varphi$ (I-II columns) and $\sigma$ (III-IV columns), for the ground truth (GT) and for different values of the iteration step $k$, in the case of the initial guess $(\varphi_0=0,\sigma_0=1)$.
• Figure 2: Plots of the initial condition and the final condition at $t=30$ days, for both the variables $\varphi$ (I-II columns) and $\sigma$ (III-IV columns), for the ground truth (GT) and for different values of the iteration step $k$, in the case in which the initial guess is the characteristic function of a circle shifted from the ground truth configuration.
• Figure 3: Plots of the optimisation functional $\mathcal{J}$, of the norm $\lVert v_k\rVert_H$ and of the learning rate $\rho^{m_k}$ vs the iteration steps $k$ for the cases ${\alpha}=0.01$ and ${\alpha}=0$.
• Figure 4: Plots of the refined mesh, the initial condition and the final condition at $t=20$ days for the ground truth (GT) and for different values of the iteration step $k$, in the case ${\alpha}=0.01$.
• Figure 5: Plots of the refined mesh, the initial condition and the final condition at $t=20$ days for the ground truth (GT) and for different values of the iteration step $k$, in the case ${\alpha}=0$.

Theorems & Definitions (32)

• Remark 3.1
• Theorem 3.2
• proof
• Theorem 3.3
• proof
• Remark 3.4
• Theorem 4.1
• proof
• Remark 4.2
• Remark 4.3