Abstract
2606
Introduction: The IMMUNE-AD study (NIH R01-AG022304) is investigating potential mechanisms by which physical activity counteracts the negative effects of the APOE e4 allele, a genetic risk factor for late onset Alzheimer’s disease. Among other parameters, the IMMUNE-AD study investigates amyloid accumulation over a two-year period based on 18F-florbetapir PET scans. Detecting statistically significant differences in amyloid accumulation rate (AAR) can be challenging, considering the random and non-random errors associated with PET imaging and analysis. PET image analysis methods were investigated to optimize the quality of AAR measurements.
Methods: 18F-florbetapir PET images were acquired and reconstructed under the ADNI protocol (4x5min frames, 50-70min post-injection) at a single site with a single scanner model (Biograph mCT 4-ring). Participants are cognitively normal elders (65-80yrs), of which 16 e4+ (e3e4) and 10 e4- (e3e3) have completed both initial PET scans and two-year follow-up PET scans to date. PET images were processed according to two methods with published Centiloid (CL) calibrations for 18F-florbetapir. The first method “GAAIN” (Navitsky et al, Alz Dement 2018) applies spatial normalization of the participant’s MRI and PET images to the MNI152 template and calculates SUVR using the standard GAAIN regions referenced to the whole cerebellum. The second method “PUP” (Su et al, Alz Dement DADM 2019) uses FreeSurfer and the PUP pipeline (https://github.com/ysu001/PUP) to generate patient-specific cortical regions and to calculate cortical SUVR with regional partial volume correction referenced to the cerebellar cortex. For both methods, AAR is calculated (in units CL/year) most simply as the difference between CL scores of the two timepoints, divided by the time between PET scans. Longitudinal versions “LPUP” and “LGAAIN” of the two pipelines were investigated, where images from the two PET timepoints were mapped to a common MRI. In addition, the average “Avg” of the LGAAIN and LPUP methods was investigated.
Results: Participants with average CL<15 and average CL>15 were deemed as “Non-Accumulators” and “Accumulators”, respectively. AAR data supported this grouping, with Non-Accumulators having AARs clustered around zero, and Accumulators having highest AARs. The ability to distinguish between the two groups is related to the AAR signal (average Accumulator AAR) and the noise (standard deviation in Non-Accumulator AAR). The ratio of signal to noise (AAR-SNR) is a figure of merit of the AAR analysis method. Although the PUP method had more consistent baseline CL scores compared to the GAAIN method, its AAR-SNR (2.6) and Non-Accumulator noise (2.6 CL/yr) were worst among the methods. The GAAIN method had higher AAR-SNR (4.3) and lower noise (1.8 CL/yr), indicating a significant advantage for longitudinal comparisons. The longitudinal methods improved both AAR-SNRs (3.2 and 4.8 for LPUP and LGAAIN) and noise (1.7 CL/yr and 1.5 CL/yr) by avoiding errors associated with use of multiple MRIs. Best results were obtained when averaging the LGAAIN and LPUP results, yielding the highest AAR-SNR (5.0) and lowest noise (1.3 CL/yr).
Conclusions: Longitudinal analysis of AAR depends on different factors than single-timepoint analysis and requires different processing methods for optimal results. Adapting processing pipelines to improve consistency (e.g., a single MRI, single spatial normalization or segmentation) is recommended. Based on this initial analysis, averaging results from substantially different methods appears to improve longitudinal results further.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.