J. Falco-Roget, A.Crimi
Graph representation learning methods have recently been applied to predict how brain functional and structural networks will evolve in time. However, to obtain minimally coherent predictions, these methods require large datasets that are rarely available in sensitive settings such as brain tumors. Because of this, the problem of plasticity reorganization after tumor resection has been largely neglected in the machine learning community despite having an enormous potential for surgical planning. We present a machine learning model able to predict brain graphs following brain surgery, which can provide valuable information to surgeons planning better surgery. We rely on the idea that surgical outcomes share network similarities with healthy subjects and combine them in a Bayesian approach. We show how our method significantly outperforms simpler models even when taking advantage of the same prior. Furthermore, generated brain graphs share topological features with the real brain graphs. Overall, we present the problem of plasticity reorganization after brain surgery in a normative manner while still achieving competitive results.