An automated clustering algorithm for reference region extraction of brain 11C-PK11195 studies

Background: A convenient way to analyze quantitatively 11C-PK11195 cerebral PET scans is the simplified reference tissue model (SRTM), which fits Time Activity Curves (TACs) to those of a reference non-pathological region. In this work, we present a fully automatic method which makes use of the expected tracer concentration in 4 predefined tissue classes (gray matter, white matter, blood, high specific binding) to automatically produce the TAC of the reference region. Methods: The proposed algorithm starts by removing irrelevant structures (skin and skull) based on activity and shape information, then it classifies the remaining voxels in the 4 tissue classes based on the L
²
distance from pre-defined seeds. Within those classified as “gray matter”, those having the greatest L
²
distance from the high specific binding, are selected to build the reference TAC. Seeds were built from scans of 9 healthy controls and from 2 patients with Alzheimer dementia, known to have activated microglia. Binding potential values (BP) were measured in two structures for 9 healthy controls using the SRTM with the output of clustering algorithm or a whole cerebellum reference as input and then compared. Results: The output maps of clusters did not show evident artifacts on visual inspection. The (BP) obtained by the proposed algorithm was close to the one obtained with the cerebellum reference in all cases. Bland-Altman analysis found a mean difference of -0.086 and 95% limits of agreement of -0.17, -0.05. Conclusions: The proposed clustering algorithm is a fully automated one, which showed a reliable performance in the healthy controls and was comparable to the reference method.