Philip De Witt Hamer

Amsterdam UMC, location Vumc

Machine learning for better neurosurgical decisions in patients with glioblastoma

The most common brain tumor - glioblastoma - is a rare cancer but invariably fatal, despite surgery and chemoradiotherapy. Prolongation of patient survival and persistence of quality of life critically depend on decisions by neurosurgeons. The aim of this study is to improve these decisions in new patients by digital support systems based on patient characteristics and standard brain scans that predict decisions from experts, the tumor's growth speed, the location of tumor recurrence and the patient’s survival time. For these predictions we will combine datasets with large numbers of patients and scans in a collaboration of medical experts and machine learning experts. This should enable decision support systems to distribute expert knowledge for better neurosurgical decisions in patients with glioblastoma.

Brief summary of results of this research project

We have observed substantial variation in the treatment of patients with glioblastoma between hospitals. To better understand the source of this variation and ultimately come to better treatment standards, large datasets with MRI scans of many patients are required. Manual delineation of the location of the tumor in these scans is a labour intensive and time-consuming process. We have therefore introduced machine learning based automatic segmentation and have shown that this is an accurate and reliable method that can be used to extract relevant tumor characteristics in scans both before and after surgery.

Using these characteristics, we have been able to gain valuable insight in which of these characteristics are important in deciding which patients are eligible for surgical treatment of the tumor. Using the automatically extracted tumor characteristics we were able to more accurately predict the decision to resect or biopsy the tumor than using only the clinical information. Our next goal is to also predict the survival time, initial results are promising, and we aim to further improve these predictions by applying novel machine-learning approaches.