Magia: Robust automated image processing and kinetic modeling toolbox for PET neuroinformatics

Introduction Modelling of the radioactivity images produced by PET scanners into biologically meaningful quantities, such as binding potential, is a complex multi-stage process involving data retrieval, preprocessing, drawing reference regions, kinetic modelling, and post-processing of parametric images. The process is challenging to automatize mainly because of manual work related to input generation, thus prohibiting large-scale standardized analysis of brain PET data. To resolve this problem, we introduce the Magia pipeline that enables processing of brain PET data with minimal user intervention. We investigated the accuracy of Magia in the automatic brain-PET data processing with four tracers binding to different binding sites: [₁₁C]raclopride, [₁₁C]carfentanil, [₁₁C]MADAM, and [₁₁C]PiB.

Materials and methods For each tracer, we processed 30 historical control subjects’ data with manual and automated methods. Five persons manually delineated the reference regions (cerebellar or occipital cortex depending on tracer) for each subject according to written and visual instructions. The automatic reference-region extraction was based on FreeSurfer parcellations. We first assessed inter-operator variance resulting from manual delineation of reference regions. Then we compared the differences between the manually and automatically produced reference regions and the subsequently obtained metrics.

Results The manually delineated reference regions were remarkably different from each other. The differences translated into differences in outcome measures (binding potential or SUV-ratio), and the intra-class correlation coefficients were between 47 % and 96 % for the tracers. While the Magia-derived reference regions were topographically very different from the manually defined reference regions, Magia produced outcome measures highly consistent with average of the manually obtained estimates. For [₁₁C]carfentanil and [₁₁C]PiB there was no bias, while for [₁₁C]raclopride and [₁₁C]MADAM Magia produced 3-5 % higher binding potentials as a result of slightly lower time-integrals of reference region time-activity curves.

Conclusion Even if Magia produces reference regions that are anatomically different from manually drawn reference regions, the resulting outcome measures are highly similar. Based on these results and considering the high inter-operator variance of the manual method, the high level of standardization and strong scalability of Magia, we conclude that Magia can be reliably used to process brain PET data.