Evaluation of a Sensitive Visual Read Algorithm for Assessing 3R/4R Tau PET Images

Abstract

Developing and validating sensitive visual read algorithms for assessing Alzheimer disease–related tau in tau PET imaging is imperative, considering the implementation of the methodology in clinical practice and trials. Our aim was to compare 2 visual read algorithms for tau PET images to semiquantitative measurements and plasma phosphotau 217 (p-tau217) status. Methods: In total, 1,654 participants were consecutively recruited from secondary memory clinics in southern Sweden as part of the prospective BioFINDER-2 cohort study (May 2017–September 2023). All participants underwent [18F]RO948 scans, and 37 participants underwent an additional [18F]flortaucipir scan. PET scans were read visually in accordance with the BioFINDER visual read (BF-VR) protocol and the established visual read method for [18F]flortaucipir (FTP-VR). Comparative analyses were conducted with semiquantitative SUV ratios (SUVRs) in the entorhinal cortex (ERC) and a temporal meta-region of interest and with plasma p-tau217 status. The primary endpoints of the study included the accuracy of visual read algorithms in detecting increases in semiquantitative SUVRs and p-tau217. Secondary outcomes included the intrarater and interrater reliabilities of the BF-VR. Results: Both visual read methods exhibited strong concordance with semiquantitative SUVRs. However, the BF-VR method demonstrated superior accuracy for tau in the ERC (93.9%; 95% CI, 92.6%–95.0%) compared with the FTP-VR method (89.9%; 95% CI, 88.3%–91.3%; P, 0.0001). The BF-VRs displayed lower accuracy when detecting tau in the temporal meta-region of interest (91.8%; 95% CI, 90.4%–93.1%) compared with the FTP-VRs (94.7%; 95% CI, 93.6%–95.8%; P 5 0.002). Further, the BF-VRs exhibited similar accuracy (0.866; 95% CI, 0.847–0.884) to the FTP-VRs (0.857; 95% CI, 0.837–0.875; P 5 0.54) for detection of p-tau217 abnormality. The interrater reliability of the BF-VR algorithm was excellent (weighted Cohen k, 0.87; range, 0.82– 0.93), and intrarater reliability was almost perfect (Cohen k, 0.94; range, 0.89–0.98). The concordance of the BF-VR assessments of [18F]RO948 and [18F]flortaucipir images was excellent (Cohen k, 0.94; range, 0.86–1.00). Conclusion: Visual reads offer a straightforward method for evaluating tau PET scans. The BF-VR algorithm provides a more accurate algorithm for detecting tau uptake in the ERC compared with the established FTP-VR algorithm and exhibits similar performance when using [18F]RO948 and [18F]flortaucipir images. The BF-VR algorithm further showed excellent interrater and intrarater reliabilities.