Abstract
3267
Introduction: PET imaging is being used to conduct drug occupancy and affinity studies as part of the drug development process. We recently introduced voxel-level images of drug occupancy created from PET images via our “Lassen plot filter” [de Laat et al., 2020]. EC50 images were developed by applying the Emax model at the voxel level to10 occupancy images [de Laat et al., 2021]. Fits of the Emax model to voxel-level data can be more variable than fits to regional or whole-brain data. As a consequence, voxel-level EC50 images can be noisy. Our purpose here is to investigate an image segmentation method for reducing noise in EC50 images so that we can estimate regional differences in drug affinity with maximal confidence.
Methods: A digital phantom was created to represent a series of 10 occupancy images, corresponding to different plasma concentrations of drug. The phantom shape was a large sphere with EC50 = 8 ng/mL, and maximum occupancy (Omax) = 1, containing 6 smaller spheres with EC50 = 8 ng/mL and Omax = 0.6. An established noise model (based on observed variability in measured occupancy values in humans) was applied to the simulated occupancy images [Hoye et al., 2021]. We fitted two versions of the Emax model (to yield EC50) to the noisy voxel-level occupancy data either by fixing Omax to 1 (‘1-parameter model’) or by simultaneously estimating Omax from the data (‘2-parameter model’). Voxel-level best-fit Omax, EC50, and coefficient of variation (CV(EC50)) images were generated. Parametric images were constructed with the parameter estimate of the best model at each voxel, using the corrected Akaike information criterion (AICc) to determine the best model fit. A previously published clustering algorithm, SLICR, [Achanta et al., 2012 and Wu et al., 2019] was modified to functionally segment occupancy images into “k” clusters. SLICR is a k-means algorithm for clustering, in which the distance calculation is reduced by limiting the search space. A hyper-parameter (“m2”, the strength of the spatial distance term) is selected to control compactness and size of the clusters. Mean occupancy of the voxels within each cluster was used to generate cluster-level best-fit (based on the AICc) Omax, EC50, and CV(EC50) images.
Results: <h4>Figure 1 shows the Omax, EC50 and CV(EC50) mid-sagittal images for the Ground Truth of the digital phantom, voxel-level fitting alone, and cluster-level fitting (SLICR). Occupancy images were divided into 1300 clusters with “m” = 1.9. Post-processing with SLICR generated EC50 images with similar accuracy as the voxel-level fits in both the large sphere (Median: 8.0, and 8.0, respectively) and in the small spheres (8.1, and 8.0, respectively), as well as in Omax images of the large sphere (1.0, and 1.0, respectively) and small spheres (0.61, and 0.61, respectively) as shown in Table 1. Notably, the values in CV(EC50) images generated by cluster fitting with SLICR were an order of magnitude smaller than those from the voxel-level fitting for the large sphere (0.01, and 0.18, respectively) and the small spheres (0.03, and 0.69, respectively). </h4>
Conclusions: EC50 images generated by fitting Emax at the cluster-level with SLICR yielded similar accuracy, but greatly improved precision compared to the voxel-level estimation. These results suggest that SLICR will provide more confidence in our estimates of regional drug affinity and could improve drug development.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.