Session Item

Modelling radiation response of radiotherapy is a complex task, and most models on clinical data only consider organ level effects. Capturing local tissue response in vivo may allow for clearer linkage of dose and biological effect. There is lack of robust methodology to incorporate longitudinal response measures, anatomical changes over time, and the complex structure of multilevel data. We proposed a framework to model radiation response in vivo, using imaging biomarkers as a surrogate, which incorporates all these aspects. The framework is available online.

We demonstrate the framework with an example cohort of 6 paediatric patients. Change of fractional anisotropy on diffusion tensor MRI images was used as imaging response biomarker as a loss of voxel intensity here has previously been linked to radiation damage. CT, PET and other MRI sequences were included in the framework to show the multi-modality capabilities.

By acquiring images before and at multiple timepoints after radiotherapy, a longitudinal dataset was collected. The images were combined into hybrid templates at each timepoint, registered internally by affine transformation. Each follow-up template was registered to the baseline template through deformable hybrid template image registration. As we included the planning CT scan, we created a dataset with per-voxel information of radiation dose and imaging change over time.

Imaging changes over time were modelled on a voxel-by-voxel level with a generalised linear mixed regression model, accounting for inter- and intra-patient variations through nesting. To demonstrate the modelling aspect of our framework, we considered biological effect of radiation dose and LET of proton radiotherapy, according to:

where n denotes the voxel. We modelled the dataset both in a combined model across all patients and follow-ups, and separately for each patient and follow-up. The latter to assess inter- and intra-patient variations.

The framework was successful in registering multimodal, longitudinal imaging data, see Figure 1. This enabled spatial mapping and modelling of radiation response.

Both our combined model and individual models per follow-up correctly reproduced the expected dose-dependency signal; however, we saw no clear dependence of LET (the c term); see Figure 2.The inter- and intra-patient variations in radio sensitivity appeared to overshadow any LET effects on tissue radiation response. These specific results are discussed further elsewhere [1]. The complete framework is available online [2].

We have demonstrated a powerful framework to rigorously model localised, in-vivo radiation response effects in longitudinal datasets, while accounting for anatomical changes over time. We demonstrated the ability to model imaging changes and reproduce the expected result. We also saw large variations between patients, indicating a need for accurate or composite, multimodal imaging biomarkers which the framework is capable of handling.