Session Item

Post-radiation morphological changes in the brain have been investigated, but remain challenging to quantify. Observed changes, such as tissue atrophy, have often been compared to chronological age and could be linked to cognitive decline. This study aims to provide an interpretable score to quantify post-radiation changes in the brain using the framework “brain age gap estimation (BrainAGE)”, which is free of any pre-engineered features, like cortical thickness or regional volumes (e.g.: hippocampus). BrainAGE utilizes deep learning models trained on healthy MRI scans to predict the biological age for patients. By obtaining ages for multiple longitudinal scans, the rate of change within the brain can be quantified with an interpretable score, represented by an aging rate.

129 longitudinal MRI scans that have been collected from 32 glioma patients who received radiotherapy at UMC Utrecht were analyzed retrospectively using a neural network-based deep learning model for brain age prediction. This model was pre-trained on 14503 healthy MRI scans from the UK Biobank. An age was predicted for each MRI scan, which was fed into a linear mixed effects model to predict aging rates for every patient. Saliency maps were extracted from the model and averaged from the whole study population in order to provide insight into which parts of the brain contribute the most to the estimated brain age. Hotspots from the saliency maps were quantified via existing MNI space-available brain atlases, like the Oxford-Harvard cortical and subcortical atlas.

Fig. 1/A shows an example of a 42-year-old female patient with an increase in the predicted brain age as time progresses, indicating radiation-induced accelerated aging. Furthermore, Fig. 1/B shows the population average saliency map, indicating which areas have relatively high contribution in estimating the biological age according to the neural network. Among the most contributing regions, numerous anatomically well-defined areas can be found such as the Heschl’s gyrus and Middle Cerebellar Peduncle. Fig. 2 shows the predicted aging rate for each individual patient and the group aging rate in red. The full cohort of patients show the mean aging rate of 2.78 years per year (95% CI = 2.54-3.02), significantly higher than the normal aging rate of 1 (p = 6.72e-16).

The BrainAGE framework shows a significantly higher rate of biological aging for glioma patients after radiotherapy. Due to the lack of pre-engineered features, BrainAGE can provide an overview of the spatial post-radiation changes in the brain, taking the entire brain volume into account. The well-defined areas obtained from the saliency maps could indicate importance for future research into post-radiotherapy changes in the brain. Applying this framework provides further insights into radiation-induced accelerated aging. Use of BrainAGE may eventually guide radiation therapy planning and contribute to expectation management of treatment side effects.