Visual Assessment of Growth Prediction in Brain Structures after Pediatric Radiotherapy

Abstract

Pediatric brain tumor radiotherapy research is investigating how radiation influences the development and function of a patient's brain. To better understand how brain growth is affected by the treatment, the brain structures of the patient need to be explored and analyzed pre- and post-treatment. In this way, anatomical changes are observed over a long period and are assessed as potential early markers of cognitive or functional damage. In this early work, we propose an automated approach for the visual assessment of the growth prediction of brain structures in pediatric brain tumor radiotherapy patients. Our approach reduces the need for re-segmentation and the time required for it. We employ as a basis pre-treatment Computed Tomography (CT) scans with manual delineations (i.e., segmentation masks) of specific brain structures of interest. These pre-treatment masks are used as initialization, to predict the corresponding masks on multiple post-treatment follow-up Magnetic Resonance (MR) images, using an active contour model approach. For the accuracy quantification of the automatically predicted posttreatment masks, a support vector regressor (SVR) with features related to geometry, intensity, and gradients is trained on the pre-treatment data. Finally, a distance transform is employed to calculate the distances between pre- and post-treatment data and to visualize the predicted growth of a brain structure, along with its respective accuracy. Although segmentations of larger structures are more accurately predicted, the growth behavior of all structures is learned correctly, as indicated by the SVR results. This suggests that our pipeline is a positive initial step for the visual assessment of brain structure growth prediction.