Introduction

A central question in the delivery of radiotherapy (RT) to high-grade gliomas is how to target tumor regions which will be resilient to radiotherapy. Advances in medical imaging, especially with multi-parametric magnetic resonance imaging (mpMRI), now enables a data-rich series of noninvasive measurements of tumor properties including cellularity, tissue structure, blood flow, and blood volume that provide insight into treatment response. It remains, however, how to leverage these data to optimize radiotherapy in anticipation of treatment resiliency. Towards this end, we have developed an approach to forecast tumor response during radiotherapy through the use of image-driven biologically-based mathematical models.   

Methods

Image acquisition—Longitudinal mpMRI data was collected from twenty patients with high-grade gliomas as part of a prospective radiotherapy trial at the MD Anderson Cancer Center. The imaging schedule included pre-RT scans, weekly scans during RT, and monthly scans after RT completion. The mpMRI protocol included T₂-fluid attenuated inversion recovery (FLAIR), T₁-weighted with or without contrast, and diffusion imaging (diffusion tensor imaging or diffusion weighted imaging). All images were registered longitudinally to the pre-RT visit. The tumor burden was expertly segmented into enhancing and non-enhancing regions using contrast-enhanced T₁-weighted and T₂-FLAIR images, respectively. Additionally, the apparent diffusion coefficient obtained from diffusion imaging was utilized to estimate the tumor cell count within the enhancing tumor region.

Computational modeling—A family of 44 biology-based models were developed characterizing the enhancing and non-enhancing components of high-grade glioma’s response to chemoradiation. This model family consists of a set of nested 3D reaction-diffusion models with parameters describing tumor cell mobility (or diffusion), proliferation (or reaction), tissue mechanics, death due to RT, and death due to chemotherapy. For each patient, data from pre-RT to week 3 of RT were used to calibrate model parameters capturing the unique, patient-specific, characteristics of growth and response. These patient-specific parameters were then used to forecast, on a patient-specific basis, response in the near term (end of RT) and longer-term (1-month or later after RT). The forecasted response from every member of the model family was combined into an ensemble prediction and compared directly to the observed response at the global (total tumor cell count; TTC) and local levels (concordance correlation coefficient; CCC).

Results

At the global level, we observed a high level of agreement between the ensemble predicted and observed total TTC at the end of RT (CCC = 0.96) and at 1-month post-RT (CCC = 0.91). Likewise at the local level, we observed a strong level of agreement of voxel-level measures of tumor cell density between the predicted and observed response at the end of RT (CCC median = 0.86; IQR = 0.16) and at 1-month (CCC median = 0.77; IQR = 0.17).

Conclusion

Image-driven mathematical modeling informed by mpMRI collected before and during radiotherapy can yield patient-specific forecasts of response with low error at the global and local levels.