Defining a Centiloid scale threshold predicting long-term progression to dementia in patients attending the memory clinic: an [¹⁸F] flutemetamol amyloid PET study

Positron emission tomography (PET) using radiotracers specific for amyloid-β(Aβ) plaques has been a breakthrough for Alzheimer’s disease (AD) research [1]. Several PET studies have demonstrated that Aβ plaques are detectable in the brain of non-demented older adults [2]; and that high Aβ-PET signal is associated with subsequent cognitive decline [3‐5] and greater risk for progression to dementia [6, 7]. The results of these research studies make Aβ-PET a promising biomarker for Memory Clinics to evaluate the risk of individual patients to develop dementia. So far, three Aβ tracers (florbetapir, florbetaben, and flutemetamol) have been approved for clinical use. Specifically, [¹⁸F] flutemetamol is approved in the USA and the European Union to visually rule out Aβ as the probable cause of the cognitive deficits. PET semi-quantitation could potentially increase sensitivity and detect Aβ earlier than visual readouts [8]; but until recently, the use of multiple radiotracers and the lack of standardization in semi-quantitative measurements have hampered the field by limiting between-study comparability. Consequently, PET quantitation is commonly used in research but is not routine in clinical practice.

The Centiloid project was designed to address this issue by proposing a standardized Aβ-PET processing pipeline and a method to transform resulting PET signal measures obtained with different radiotracers into a common unit, called “Centiloid” [9]. The Centiloid Aβ-PET scale is anchored at 0 and 100 Centiloids, with 0 Centiloid score reflecting a definitively Aβ-negative brain (originally calculated as the average value of a group of healthy subjects below the age of 45) and 100 Centiloids reflecting the average signal observed in patients with typical mild or moderate AD dementia. This harmonized method (originally anchored to [¹¹C] PiB but now applicable to [¹⁸F] tracers after calibration) has great potential to produce cohesive and comparable results from disparate clinics across the world.

The present study compared visual and semi-quantitative assessments of [¹⁸F] flutemetamol PET data in a cohort of normal older adults and non-demented patients recruited at a single Memory Clinic. We used the overall predictive value to assess progression to dementia after a long-term follow-up as our standard of truth and tested different Centiloid thresholds. As a secondary objective, we evaluated whether PET semi-quantitation could help detecting visually negative cases that would subsequently progress to visually positive in a subset of participants with longitudinal PET data.

In this study, we sought to compare the long-term predictive values of qualitative, visual assessments and quantitative, Centiloid assessments of [¹⁸F] flutemetamol Aβ-PET images acquired in a monocentric cohort of non-demented participants attending the Memory Clinic. By assessing true and false positive in amyloid PET-diagnosed patients, this prospective diagnostic accuracy study accomplishes the primary aim of the 4th phase of an AD biomarker development [19, 20]. We observed that a threshold set at Centiloid = 26 best discriminated participants who will progress to dementia from participants who will remain clinically stable 6 years after PET. Using a Centiloid = 26 threshold increased the predictive value of clinical diagnoses: Overall, 80% of the MCI and 17% of the CN/SCD progressed to dementia after 6 years. Among the Aβ + (PPV), 94% of the MCI and 62% of the CN/SCD progressed, while among the Aβ-negative (NPV), 47% of the MCI and 97% of the CN/SCD remained clinically stable.

Visual assessment provided a similar, yet marginally lower, overall predictive value compared to a quantitative threshold set at Centiloid = 26. Using positive visual assessments as Aβ-PET readouts resulted in higher specificity (94% vs. 90%) and lower sensitivity (76% vs. 83%) for subsequent dementia compared to Centiloid assessments. Of note, only eight participants (5%) had a baseline Centiloid ≥ 26 but were not classified as visually positive. Among these cases, three progressed to dementia, two remained stable for longer than 6 years, and three were only followed for 2 years and were not included for deriving predictive values (Fig. 4). Therefore, the overall predictions of quantitative (87%) and visual (86%) assessments only differed by one of the 98 participants (+ 1%, not significant). A previous study with post-mortem validation in 78 cases did not observe either different accuracy between visual and quantitative Aβ-PET assessments [21]. Larger studies are needed to further evaluate the added value of quantitative over visual assessments, particularly in discordant cases, or to increase the sensitivity of early Aβ detection. We are only aware of one large study that specifically looked into the predictive power of visual versus quantitative Aβ-PET measurements: In 401 MCI participants from the ADNI cohort [22], the authors reported that visual reads had higher specificity (96% vs. 90%) and lower sensitivity (79% vs. 85%) than the SUVr values, with similar overall predictive power. However, follow-up length was relatively short (1.6 years), preventing a valid comparison of the utility of quantitation and visual reads for predicting how discordant cases evolve over a longer follow-up period. In a longer study (up to 3 years) [5], flutemetamol PET images were found positive in 52 of 81 amnestic MCI who converted to AD dementia (sensitivity, 64%), while they were negative in 99 of 143 non-converters (specificity, 69%); however, this study did not compare quantitation and visual assessments.

Importantly, we observed that the predictive value of a positive scan significantly dropped from 88 to 67% when including the participants who were followed for short periods (< 4 years), highlighting the importance of follow-up duration for establishing a clinical standard-of-truth. Of note, the positive predictive value is excluded from the routine approved label for Aβ-PET imaging (https://​www.​ema.​europa.​eu/​en/​documents/​product-information/​vizamyl-epar-product-information_​en.​pdf) since studies used for registration purposes were deemed to have too short a follow-up period for this claim to be made. Studies therefore as the one described in this paper with a longer term follow-up could contribute to the body of data highlighting the risk of clinical progression in the presence of Aβ pathology, that is, according to our data, 67% risk of progressing to dementia after 4 years, and 88% after 6 years.

We also observed lower PPV when using a Centiloid = 12 threshold, derived from Aβ assessment at post-mortem [18] (moderate to frequent CERAD plaque count). It is possible that participants with early Aβ, as indicated by a Centiloid in the 12–25 range, will progress to dementia many years after PET and that our clinical endpoint did not capture the slowly developing preclinical pathological process. Therefore, longer studies, up to a decade or more, are needed to accurately evaluate the impact of early Aβ burden on subsequent progression to dementia [23, 24]. Nevertheless, our longitudinal PET study did not observe any change in participants with sub-threshold signal, nor did the larger Mayo Clinic Study of Aging (Centiloid threshold = 19) [25], suggesting that an early Aβ-PET signal close to Centiloid = 12 is hardly distinguishable from noise. Interestingly, in this same PET autopsy study [18] as well as in a previous report [26], a higher threshold (Centiloid = 24) was observed against intermediate to high levels of AD neuropathological changes including thus both Aβ and tau lesions as a gold standard (unlike CERAD plaque count), which may be closer to a dementia endpoint. A recent paper observed that a Centiloid threshold = 21 detected moderate or frequent plaque density while Centiloid = 10 was optimal for excluding neuritic plaques [27]. They also observed higher threshold (Centiloid = 49) against intermediate to high levels of AD neuropathological changes. Consistently in CSF studies, PET thresholds obtained against CSF Aβ₄₂ are low, at Centiloid = 12, while PET thresholds obtained against CSF total tau, phosphorylated tau, or tau/Aβ₄₂ ratio are in the range of 25–33 [28]. Several studies using various endpoints (subsequent dementia, change in PET signal, CSF, histopathology [29]) generally converge around a Centiloid threshold = 26 (± 7), which appears to discriminate most accurately between those individuals on an AD trajectory from those who are not. Lower cut-off values can be used to detect incipient Aβ pathology, but these are not associated with progression to AD dementia, even after a median follow-up of 6 years.

From a clinical imaging standpoint, previous studies using visual ratings with [¹¹C] PiB [30], [¹⁸F] flutemetamol [31], [¹⁸F] florbetaben [21], and [¹⁸F] florbetapir [32] showed good inter-rater reliability, especially among experienced readers. With the addition of quantitative information to visual reads, most raters change from “non-elevated” visual reads to “elevated” reads [8]. For experienced readers, the improvement in accuracy and confidence is likely restricted to a few borderline, but clinically relevant, cases [32] indicating that for the majority of images there is high concordance between visual interpretation and quantitative analysis. We only have one experienced nuclear physician at our center to perform visual reads, preventing us from evaluating inter-rater reliability. Still, we can confirm that using Centiloid = 26 as a threshold reclassified all our visually borderline cases (100%) and five of 91 (5.5%) negative visual reads to positive, while no visually positive cases were reclassified. Using Centiloid = 26 as a threshold also allowed detection of visually negative cases who progressed to visually positive at a subsequent scan, on average 3 years later. We therefore see the potential for the use of semi-quantitative assessment, either SUVr or a standardized metric such as the Centiloid scale, for increasing the sensitivity and the reliability over time of Aβ-PET interpretation in non-demented patients attending Memory Clinics. Future work will further investigate how regional PET signal—assessed visually or using quantitation—could contribute to dementia prediction [6, 33].

In conclusion, our results indicate that augmentation of visual interpretation of [¹⁸F]-flutemetamol Aβ-PET images with semi-quantitative, SUVr, or Centiloid information improves the sensitivity of visual assessment in some negative, and visually borderline, cases. While SUVr may be more readily available, the Centiloid scale offers easier comparison with other centers. When semi-quantitation is not available, we recommend considering visually borderline cases as positive. As visual and quantitative assessments provide the same classification for most patients, larger studies would be needed to demonstrate statistical differences between those assessments.