Evaluating robustness of the Centiloid scale against variations in amyloid PET image resolution

Background The Centiloid method enables direct quantitative comparison of amyloid‐PET scans acquired with different amyloid tracers. However, it does not explicitly model differences in image resolution which may have an impact on quantification, particularly in multicenter studies. This study evaluated the stability of the Centiloid scale with or without post‐processing smoothing to simulate distinct reconstruction settings. In addition, the effect of reference region selection was assessed. Methods A total of 318 [ 18 F]flutemetamol PET scans from the ADC and EMIF‐AD Twin 60++ cohorts are included in this study (Table 1). Scans were reconstructed with LOR‐Ramla algorithm. Additionally, images were smoothed to get an effective resolution of 8mm. All images were quantified with a validated Centiloid (CL) pipeline using the GAAIN cortical composite and five different reference regions (whole cerebellum, cerebellum gray, pons, whole cerebellum+brainstem and eroded white‐matter). To assess the impact of reconstruction settings on CL values, we calculated the mean absolute difference between CL values of the original reconstructions vs the smoothed images. Differences above 5CL were considered as relevant. Finally, Bland‐Altman plots were used to assess whether any proportional bias was present between these outcome measures. Results Figure 1 displays image quality differences between local reconstruction settings and post‐smoothed images. CL values were not significantly affected by reconstruction methods using either whole cerebellum, whole cerebellum+brainstem, or white‐matter as reference region (Table 2). However, for the images with a higher level of amyloid, ΔCL was higher than 5 using whole cerebellum. On the contrary, the CL was highly affected by image resolution when using the pons as reference region, introducing constant bias for all ranges of amyloid burden. Using cerebellum gray, ΔCL has an inverse linear relationship with amyloid burden, though it was below 5CL for subjects with low to moderate amyloid burden, and differences above 5 were observed for images with CL above 80 (Figure 2). Conclusion Whole cerebellum and whole cerebellum+brainstem, the standard reference regions in CL quantification, are capable of providing stable CL values irrespective of image resolution for a wide range of amyloid burden. Other reference regions were less stable.