Session Item

Clinical target volume (CTV) definition for adjuvant radiotherapy of glioblastomas (GBM) is commonly performed via isotropic expansion of the resection cavity and the residual gross tumor volume (GTV) or the FLAIR hyperintense region following international guidelines. A commonly used isotropic expansion of 2 centimeters is loosely based on postmortem analyses and imaging studies demonstrating that GBM recurrences occur within this distance in approximately 70-90% of cases (1–8). The efficacy of adjuvant radiotherapy proves that this approach covers a significant part of areas with microscopic tumor infiltration. On a patient-specific basis, one study showed that expansions ranging from one to three centimeters would have covered all individual locoregional recurrences. GBM growth, however, appears to be anisotropic as it is confined by anatomical barriers and happens with differing growth speeds in white and grey matter. Spatial localization of areas with microscopic infiltrations could thus be used to generate more individualized CTVs. This talk summarizes two promising solutions for the improved definition of a microscopic tumor volume using medical imaging studies.

First, mathematical tumor growth models allow predicting the path of infiltrative tumor cells. Adaptation to GBM, the microscopic spread has been modeled using partial differential equations of the reaction-diffusion type combined with information of anatomical barriers, grey and white matter, and white matter tracts. The frequently used Fisher–Kolmogorov equation describes tumor growth via the two factors “diffusion” and “proliferation”. Early models had to rely on a seed point for initiation rather than the actual GTV and used approximate values for relevant parameters such as the cell density. Recent developments incorporated multimodal imaging such as MRI and 18F-fluoro-ethyl-tyrosine (18F-FET) PET for spatially resolved tumor density estimation and patient-specific Diffusion Tensor Imaging (DTI) for the anatomic definition of anisotropic growth along white matter tracks.

Second, the development of imaging-specific artificial intelligence techniques such as deep convolution neural networks, often referred to by the term “Deep Learning”, have opened new possibilities in image processing and analysis. Deep Learning-based elimination of the free water component in peri-tumoral edema has been shown to improve the predictive value of the local recurrences based on fractional anisotropy. In another study, a convolutional neural network could be successfully trained to differentiate glioblastomas from brain metastases based on peri-tumoral free water-corrected fractional anisotropy.Moreover, machine learning algorithms have been deployed for tumor growth modeling to allow training on patient-specific data. Through the integration of such techniques with mechanistic models growth prediction could further be improved. Alternatively, generative adversarial networks or convolutional neural networks were directly trained to predict tumor growth based on longitudinal imaging studies. 

In conclusion, Deep Learning and tumor growth modeling-based estimation of microscopic invasion of GBM may be used for individualized CTV definitions. In patients with a high risk for more distant progression, larger anisotropic CTVs may be indicated. In patients with a low risk for distant progression, smaller CTVs may be sufficient which could lead to reduced toxicity and improved opportunities for re-irradiations. The improvement of contemporary techniques and their validation in future prospective trials is warranted before clinical application.