Abstract
INTRODUCTION: Brain and central nervous system (CNS) tumors are the second most common cancer type in children and adolescents. Positron emission tomography (PET) imaging with radiolabeled amino acids visualizes the amino acid uptake in brain tumor cells compared with the healthy brain tissue, which provides additional information over magnetic resonance imaging (MRI) for differential diagnosis, treatment planning, and the differentiation of tumor relapse from treatment-related changes. However, tumor delineation is a time-consuming task subject to inter-rater variability. We propose a deep learning method for the automatic delineation of O-(2-[(18)F]fluoroethyl)-l-tyrosine ([(18)F]FET PET) pediatric CNS tumors. METHODS: A total of 109 [(18)F]FET PET and MRI scans from 66 pediatric patients with manually delineated reference were included. We trained an artificial neural network (ANN) for automatic delineation and compared its performance against the manual reference on delineation accuracy and subsequent clinical metric accuracy. For clinical metrics, we extracted the biological tumor volume (BTV) and tumor-to-background mean and max (TBR(mean) and TBR(max)). RESULTS: The ANN produced high tumor overlap (median dice-similarity coefficient [DSC] of 0.93). The clinical metrics extracted with the manual reference and the ANN were highly correlated (r ≥ 0.99). The spatial location of TBR(max) was identical in almost all cases (96%). The ANN and the manual reference produced similar changes in the clinical metrics between baseline and follow-up scans. CONCLUSION: The proposed ANN achieved high concordance with the manual reference and may be an important tool for decision aid, limiting inter-reader variance and improving longitudinal evaluation in clinical routine, and for future multicenter studies of pediatric CNS tumors.