ESTRO 2024 - Abstract Book

S5020

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Multiparametric Magnetic Resonance Imaging (MRI) is the standard imaging modality for Glioblastoma Multiforme (GBM) and is commonly employed for diagnosis, surgical guidance, radiotherapy planning, and follow-up assessment. However, the limited resolution of current MRI modalities poses significant challenges in determining the microscopic tumor spread in the surrounding tissues beyond the visible border of Gross Target Volume (GTV) [1]. Currently, the Clinical Target Volume (CTV) is defined by adding an isotropic margin around the GTV trimmed by the anatomical barriers for GBM cells invasion into the normal tissue, although the potential limitations of this definition with respect to the anisotropic spread of the GBM cells outside the GTV are often brought to discussion. As the high recurrence rate of treated GBMs may be associated with the presence of cells infiltrated beyond the contour of the isotropically defined CTV, determining the infiltrative cell distribution outside the GTV could be used to guide the treatment and eventually improve the outcome. A recently developed framework based on a well-established biophysical model [2] was therefore implemented to simulate the invasion of GBM infiltrative cells in the brain. This study aims to present the comparison of the performance of radiomics and deep learning methods for OS time prediction based on the results of the tumor invasion model.

Figure 1. Illustration of the methodology: Simulated clinical targets obtained from the tumor invasion model were employed in radiomics and deep learning pipelines to predict the treatment outcome.

Material/Methods:

The studied dataset included 126 high-grade glioma subjects from the International Brain Tumor Segmentation 2020 (BraTS) challenge [3]. To model the invasion of tumoral cells, the PDE-based Fisher-Kolmogorov equation was employed to capture the spatiotemporal evolution of cancerous regions [4]. In this simulation, the pattern of invasion depends on the model parameters, including the diffusion-to-proliferation ratio representing the diffusability of tumor cells into normal tissues, and the ratio of diffusion coefficients in white to gray matter, which characterizes the anisotropic evolution of tumors. A range of potential values for each of these two parameters was considered.