Application of Deep Learning for standardized delineation of healthy reference regions in O-(2-[18F]fluoroethyl)-L-tyrosine (FET) PET

FET PET is a valuable tool for managing brain tumors. Quantitative image analysis is of obvious relevance in this context. One clinically useful image derived measure is the maximal tumor-to-background SUV ratio which can be used for therapy response assessment as well
as for discrimination between tumor recurrence and treatment-related changes. Computation of this ratio requires determination of the background SUV (bSUV) from a suitable ROI defined within healthy brain tissue. Currently, the standard procedure requires manual definition of the background ROI by an experienced human observer while adhering to a set of somewhat loosely defined rules. This process is time consuming and prone to inter- and intra-observer variability. The goal of this study, therefore, was development of a reliable automated method for bSUV derivation in FET PET of brain tumor patients.

Automated delineation of the healthy brain regions was performed with a residual 3D U-Net convolutional neural network (CNN). 561 FET PET scans were used for network training (N=448) and testing (N=113). In these data, reference brain regions were manually delineated by an experienced observer. The network was trained to reproduce the corresponding manual bSUVs by identifying a suitable brain ROI (rather than aiming at reproducing the manual ROI delineation). Performance of the trained network model was assessed in the test data using the fractional difference between automatically and manually derived bSUVs.

The trained U-Net was able to accurately reproduce the manually derived bSUVs in the test data: the fractional bSUV difference was (mean +/- SD)=(-0.9 +/- 5.3)% with a 95% confidence interval of [-10.9, 8.4]%.

The achieved concordance of the network's results with the given ground truth bSUV is in line with typical achievable levels of inter- and intra-observer concordance for this task. It thus might be considered for supervised routine use to reduce user workload and improve reproducibility.