Abstract
71
Objectives: Innovations in solid-state based detection technology have brought the first total-body PET scanners into fruition. In the near future, similar technology may be used to build the next-generation NeuroEXPLORER (NX) brain scanner, offering an order-of-magnitude improvement in detection sensitivity compared to the current state-of-the-art Siemen’s HRRT. We explored how this increased sensitivity could be used in conjunction with the linear parametric neurotransmitter PET (lp-ntPET) model1-4 to better detect and classify stimulus-induced neurotransmitter (NT) events. The goal of this work was to use simulations of high-count data in the striatum to 1) predict the detection sensitivity to transient NT signals of various temporal profiles, and 2) classify NT signals based on their magnitude and timing.
Methods: Striatal 11C-Raclopride (RAC) PET data were simulated in the presence of varying endogenous dopamine (DA) signals, with the extended compartmental model5. DA signals were modeled with start times between 35-44 min, peak times between 40-65 min, and amplitudes between 50-600% above baseline concentration. For each DA signal, 1000 noisy TACs were generated for the HRRT and NX. Because the NX is anticipated to be ten times as sensitive as the HRRT, the measurement variance for the NX was scaled to 10% that of the HRRT. All TACs were fitted with lp-ntPET. The detection sensitivity to each shape of DA signal was calculated. Different classification thresholds were evaluated for their accuracy in separating “early”- vs. “late”-peaking events, and “low”- vs. “high”-amplitude events. A 4D anatomical phantom of the striatum consisting of 1850 was constructed with the TPM neuromorphometrics atlas. Each voxel was simulated with either a null signal, or a positive signal classified as high/low and early/late. Variability was introduced into DA signal parameters. A weighted k-nearest neighbors (wkNN) algorithm was developed that incorporates the 6-, 18- , and 26-neighborhood of the current voxel to reclassify it, considering the original classification of its neighbors. The weight of each neighborhood was chosen to optimize sensitivity in the NX. The certainty of a classification was given by the weighted percentage of votes for the final classification. wkNN was applied for both detection and classification, to identify voxels as either null or positive, and among the positive voxels to determine if they were early/late and high/low. The final detection sensitivity and classification accuracies were calculated in the phantom for both HRRT and NX.
Results: Detection sensitivity in both scanners was higher for higher-amplitude and later-peaking signals. The NX would expand the range of signals with >80% detection sensitivity to signals with amplitudes as low as 200% above baseline and peaking as early as 6 min post-stimulus (Fig. 1A), compared to the minimum of 550% above baseline and 20 min post-stimulus peak necessary in the HRRT (Fig. 1B). In the HRRT, detection sensitivity was 31.4% and specificity was 88.9%. The NX would increase detection sensitivity and specificity to 72.3% and 92.1%, respectively. If wkNN segmentation was applied, detection sensitivity in the NX phantom increased to 91.5%, with 100% specificity. Classification of detected signals with high-amplitude and early-peaks had an excellent positive predictive value (PPV) of >97%. Low-amplitude signals had the lowest PPV of 80.7%. All classifications had a negative predictive value (NPV) of > 85%. Voxels with incorrect classifications and voxels along the borders between different classifications had the lowest levels of certainty (Fig. 1C).
Conclusions: An ultra-high sensitivity NX scanner would greatly expand the range of detectable DA signals in the striatum, compared to the current state-of-the-art. Higher detection sensitivity and a novel classification algorithm will make it possible to the accurately classify DA signals based on their amplitudes and timing.