An interpreter sits in a room where an important meeting is taking place, but street noise keeps him from catching half the words. He conveys the meaning with errors, and the longer the conversation goes on, the more distortion accumulates. That is roughly how the standard SUVR measure works in positron emission tomography (PET) with Pittsburgh Compound-B: it estimates the amount of amyloid plaques in the brain but ignores the fact that the radiotracer washes out of tissues at different rates. Amyloid PET is a way to see plaques associated with Alzheimer's disease using a substance that accumulates in them and lights up on the scan.

The gold standard for accuracy is the DVR measure, which requires an hour of dynamic recording and complex modeling. In practice, the simplified SUVR over a short window is used more often, but it overestimates burden, and the error depends on the window time and disease stage. Earlier correction methods worked only under equilibrium, which for Pittsburgh Compound-B is reached late, and therefore they could not shorten the protocol. Researchers from the Harvard Aging Brain Study solved this problem by adapting a recently proposed generalized SUVR correction that does not require equilibrium.

The scientists used data from 88 older adults, divided into three groups by amyloid burden level: low, intermediate, and high. For each group they built an empirical model linking the rate of tracer washout from the target region to its SUVR. In the low-amyloid group, the rate was taken to equal the rate in the cerebellum (the reference region) plus a constant. In the intermediate group, a piecewise linear function of the inverse SUVR was used. In the high-amyloid group, a power-law model with two parameters. Model parameters were estimated on a training subsample (one third of the data) and tested on an independent validation subsample. Then, using a correction formula that accounts for washout rates in the target and reference regions, they computed the corrected SUVRc for two time windows: 30-45 and 45-60 minutes after injection. The results were compared with the gold-standard DVR obtained from 60-minute dynamic data.

The mean bias of SUVR relative to DVR ranged from 4.8% to 14.7% depending on the group and time window, and it grew as the window increased. After correction, the bias of SUVRc dropped to a range from −0.15% to 2.2%, and the variability of the bias decreased by a factor of 1.5 to 2. In the high-amyloid group, the reduction in bias was statistically significant for 8 of 11 regions at the 30-45 minute window and for all regions at the 45-60 minute window. Lin's concordance coefficient between SUVRc and DVR rose from 0.954 to 0.987 (30-45 min window) and from 0.918 to 0.989 (45-60 min window). In longitudinal cohorts, SUVRc reflected changes in amyloid burden more accurately: for small changes (5-10%), SUVR overestimated the increase, while SUVRc gave values close to DVR; for large changes (30-70%), SUVR underestimated, while SUVRc almost matched DVR.

This is the first time SUVR correction has been applied to Pittsburgh Compound-B and validated on independent data with gold-standard DVR. The method cuts the required scan time to 15 minutes without loss of accuracy, which matters for clinical practice and large trials. Unlike neighboring studies that use machine learning for image processing or predict events from indirect signs, this work relies on kinetic principles and offers simple parametric models that can be transferred to other cohorts. No applications are yet in sight beyond research, but in the clinic this could speed up diagnosis and make it more accessible.

Take the whole work

Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.

the paper's page