Subtracting First Principal Component May Improve 4R Tau Detectability on [¹⁸F]Flortaucipir Tau PET

Visual Abstract

Abstract

[¹⁸F]flortaucipir PET imaging has shown relatively low sensitivity for detecting four-repeat (4R) tau, which limits its utility in 4R tauopathies such as progressive supranuclear palsy (PSP). Previous studies have suggested that the first component of principal-component analysis includes non–disease-related uptake in non-Alzheimer tauopathies. In this study, we tested the hypothesis that subtracting this component could increase sensitivity to 4R tau using a large PSP cohort. Methods: Prospectively recruited patients with PSP (n = 141; 80 Richardson syndrome, 44 PSP-subcortical, and 17 PSP-cortical variants) and controls (n = 102) underwent [¹⁸F]flortaucipir PET. The first principal component (PC1) was extracted from the SUV ratio (SUVR) images of 62 controls and then subtracted from the remaining controls (n = 40) and all PSP patients. We compared the diagnostic performance before and after PC1 subtraction by evaluating SUVR differences between PSP and controls, as well as across PSP clinical variants, and by calculating correlation coefficients between SUVR and the severity of tau pathology. Results: Area under the receiver operating characteristic curves differentiating PSP and controls were improved by PC1 subtraction in the frontal white matter (0.54 to 0.67; P = 0.002) and subcortical (0.69 to 0.87; P < 0.001) regions. In the frontal regions, no differences were observed across PSP clinical variants before PC1 subtraction, whereas the PSP-cortical variant showed an SUVR higher than that of the PSP Richardson syndrome and PSP-subcortical variants after PC1 subtraction (adjusted P < 0.05). A correlation with the severity of tau pathology was observed only in the red nucleus before PC1 subtraction, whereas correlations were observed in the precentral, superior frontal gyrus, and red nucleus after PC1 subtraction (adjusted P < 0.05). Conclusion: Our approach enhanced the diagnostic performance of [¹⁸F]flortaucipir PET in PSP and increased sensitivity to 4R tau, suggesting that subtracting the first principal component improves the detectability of subtle 4R tau uptake and the potential utility of [¹⁸F]flortaucipir PET imaging for non-Alzheimer tauopathies.

The first PET tracer approved by the U.S. Food and Drug Administration for evaluating cerebral tau uptake in Alzheimer disease (AD) was [¹⁸F]flortaucipir (also known as AV-1451 or Tauvid). [¹⁸F]flortaucipir has a high affinity for the mixed three-repeat and four-repeat (4R) tau (e.g., AD tau) (1), and its uptake pattern corresponds with the accumulation of neurofibrillary tangles and high levels of AD neuropathologic changes (2). Ample evidence establishes [¹⁸F]flortaucipir as a reliable tracer for assessing tau in patients with AD. In contrast, the utility of [¹⁸F]flortaucipir PET imaging for non-AD tauopathies, such as progressive supranuclear palsy (PSP), is currently limited. Autoradiographic studies have revealed that [¹⁸F]flortaucipir has no or little affinity for 4R tau (e.g., PSP tau) (1), and there is no strong relationship between antemortem [¹⁸F]flortaucipir uptake and severity of tau pathology in PSP (3). Although [¹⁸F]flortaucipir uptake has consistently been detected in the midbrain, basal ganglia, thalamus, and cerebellar dentate in PSP (4,5), which generally aligns with the pathologic tau patterns (6), uptake in these subcortical structures included off-target signals from neuromelanin, monoamine oxidase, and iron (1,7,8). Furthermore, there is limited evidence of cortical [¹⁸F]flortaucipir uptake in patients with PSP (4) despite PSP affecting the cortex, including the frontal lobe (5).

Several studies have applied principal component analysis (PCA) or related techniques to [¹⁸F]flortaucipir PET imaging to evaluate disease-specific patterns and their correlation with disease severity. Studies of patients with AD have revealed that the first few principal components represent disease-specific patterns in both typical and atypical cases, and that these components are associated with diagnoses or disease severity (9,10). In contrast, studies of non-AD tauopathies, such as PSP (11) and frontotemporal dementia (12), consistently show that the first principal component (PC1) exhibits global uptake patterns unrelated to diagnosis or disease severity, whereas lower components (such as the second component) are related to them. These results suggest that [¹⁸F]flortaucipir uptake contains disease-related components, even in non-AD tauopathies; however, these components are relatively small and obscured by the highest variable and nonspecific global uptake patterns represented by the first component. In other words, it could be hypothesized that subtracting PC1 increases the detectability of disease-related tau uptake in patients with non-AD tauopathies.

This study aimed to test the hypothesis that subtracting PC1 could improve PSP diagnosis and increase sensitivity to 4R tau. We developed a postprocessing method to estimate and subtract PC1 from unseen [¹⁸F]flortaucipir PET images. We then compared 2 images—1 before and 1 after PC1 subtraction—in a large PSP cohort by evaluating the ability to differentiate patients with PSP and healthy controls, differentiate across PSP clinical variants, and the relationship between [¹⁸F]flortaucipir uptake and tau pathology severity at autopsy.

MATERIALS AND METHODS

Participants

The 141 patients diagnosed with PSP (13) were prospectively recruited by the Mayo Clinic, Neurodegenerative Research Group between February 2015 and January 2025. They all underwent detailed neurologic clinical evaluations, 3-T MRI, and [¹⁸F]flortaucipir PET scans. The 141 patients included 80 with PSP Richardson syndrome (PSP-RS), 44 with PSP-subcortical (24 PSP-parkinsonism, 14 PSP-progressive gait freezing, 4 PSP-postural instability, 1 PSP-ocular motor, and 1 PSP-cerebellum), and 17 with PSP-cortical (8 PSP-corticobasal syndrome, 5 PSP-speech/language, and 4 PSP-frontal) clinical variants (13). The neurologic clinical tests included the Montreal Cognitive Assessment (MoCA), Frontal Assessment Battery, Movement Disorders Society–sponsored revision of the Unified Parkinson Disease Rating Scale Part III, PSP rating scale, and PSP saccadic impairment scale (7). Of the 141 patients with PSPs, 74 underwent second scans 1 y later. Thirty-eight patients died and underwent an autopsy. Pathologic diagnoses included 32 with 4R tauopathy including PSP pathology (n = 23), atypical PSP with a cortical distribution of pathology (n = 3), and corticobasal degeneration (CBD) (n = 6), as well as multiple system atrophy (MSA) (n = 4) and other pathologies (n = 2).

In total, 102 cognitively and motorically normal participants, with a MoCA of 23 or larger (14) and a Hoehn and Yahr scale of 0, were recruited by the Neurodegenerative Research Group over the same period and underwent the same MRI and PET scans. The controls were divided into 2 groups based on age: a young group (under 65 y) and an older group (over 65 y). The younger group (n = 62) was used as an independent training cohort to extract the basis image of PC1, as described later. The older group (n = 40) was used as the age-matched cohort for evaluation. Of the 40 older controls, 23 underwent second scans 1 y later.

This study was approved by the Mayo Clinic Institutional Review Board, and informed consent was obtained from all participants.

Neuroimaging Methods

All participants underwent a standardized MRI protocol on a 3-T scanner (Magnetom Prisma; Siemens Healthineers), which included a 3-dimensional magnetization-prepared rapid acquisition gradient echo sequence. All [¹⁸F]flortaucipir PET scans were acquired on the PET/CT scanners (GE or Siemens) with approximately 370 MBq of [¹⁸F]flortaucipir. The 20-min acquisition, 80 min after the injection, consisted of 4 5-min dynamic scans. The 4-frame PET images were aligned, averaged, and rigidly registered to the magnetization-prepared rapid acquisition gradient echo image. The SUV ratio (SUVR) images were then calculated using median values in the gray matter (GM) cerebellar crus. The detailed protocols were described previously (7).

First Component Subtraction Algorithm

PCA decomposes a set of images X (=[x₁, x₂, …, x_N]) into a set of basis vectors S (=[s⁽¹⁾, s⁽²⁾, …]) and corresponding coefficient matrix H:

$$\mathbf{X}\approx \mathbf{SH}+\mathbf{\mu }{1}^T,$$

$$\mathbf{\mu }=\frac{1}{N}{\sum }_{i=1}^N{\mathit{x}}_i,$$

where N is the number of subjects and μ is the mean vector of all images. The unseen image (x_new) can be projected onto the subspace of the PC1 image using the corresponding basis vector (s⁽¹⁾) and the mean vector (μ) as follows:

$$k_1={\mathit{s}}^{\left(1\right)T}\left({\mathit{x}}_{\mathrm{new}}-\mathbf{\mu }\right),$$

$${\mathit{x}}_{\mathrm{new}}^{\left(1\right)}=k_1{\mathit{s}}^{\left(1\right)}.$$

Finally, we can obtain a subtracted image (x_sub):

$${\mathit{x}}_{\mathrm{sub}}{={\mathit{x}}_{\mathrm{new}}-\mathit{x}}_{\mathrm{new}}^{\left(1\right)}.$$

In general, PCA at the voxel level should be performed in the same space (e.g., template space) because this analysis requires all participants to have the same spatial coordinates. Ideally, however, the subtraction would be applied in the subject (native) space to maintain the original image resolution and spatial information, which are necessary for visual assessment and accurate regional analysis. Therefore, we have developed the following hybrid (template and subject) space algorithm.

Figure 1A shows the proposed flow to subtract the PC1 image in the subject space. First, the SUVR image was spatially normalized into the template space and smoothed with a 3-dimensional gaussian kernel (full width at half maximum of 3 mm) to mitigate the noise and residual normalization error. Then, the external tissue mask was applied to the image to exclude signals from nonbrain areas and converted to a vector (x). The first component vector (s⁽¹⁾) and mean vector (μ), which were precalculated from the independent cohort, were applied to calculate the coefficient (k₁) for the PC1. The PC1 image (x⁽¹⁾) was then calculated by multiplying this coefficient by the extrapolated first component image (s⁽¹⁾_extra) and then inversely warped into the subject space. The extrapolated image was precalculated using a small sphere kernel (Supplemental Fig. 1; supplemental materials are available at http://jnm.snmjournals.org) to expand the PC1 image from the tissue region to the intracranial region. This process is necessary to mitigate errors at the tissue boundary. Finally, this PC1 image was subtracted from the original image in the subject space.

FIGURE 1.

Algorithm to subtract PC1. (A) Process flow to estimate and subtract PC1. (B) PC1 images calculated from independent cohort. TIV = total intracranial volume.

We applied this algorithm to all patients with PSP (n = 141) and older controls (n = 40) for evaluation, using the data from younger controls (n = 62) as the independent cohort. The image of the PC1 (s¹) (Fig. 1B) showed a global uptake pattern, with relatively higher uptake in the subcortical nuclei and white matter (WM) and lower uptake in the GM cerebellum (including reference region). This component explained 41% of the total variance in the independent cohort.

Regional and Voxel-Based Analysis

Two types of SUVR (before and after PC1 subtraction) were evaluated in PSP-related cortical and subcortical regions. The cortical regions included the precentral, supplementary motor area, and superior frontal gyrus. The subcortical regions included the pallidum, subthalamic nucleus, thalamus, substantia nigra, red nucleus, and cerebellar dentate. These regions were selected on the basis of previous neuroimaging studies in PSP (5,7,15). The GM and WM masks were applied separately in the 3 cortical regions to evaluate the GM and WM separately. The sum of GM and WM matter masks was used in the 6 subcortical regions. In total, we evaluated 12 regions (3 GM cortical, 3 WM cortical, and 6 subcortical regions). We also combined these regions to define 3 composite areas, including GM cortical, WM cortical, and subcortical areas. The detailed procedure for the regional analysis was previously described (7).

The SUVR images were also spatially normalized into the template space and smoothed using a gaussian kernel (full width at half maximum of 6 mm) for voxel-based analysis.

Histopathologic Analysis

As described in a previous study (3), an experienced neuropathologist semiquantitatively scored the severity of tau pathology on pTau-stained slides (CP13 antibody). Severity was scored on a 4-point scale (0–3) for (1) neurofibrillary tangles, (2) neuropil threads, (3) coiled bodies, and (4) tufted astrocytes or astrocytic plaques. Evaluations were conducted in the precentral (motor) cortex, superior frontal cortex, globus pallidus, ventral thalamus, subthalamic nucleus, substantia nigra, red nucleus, and cerebellar dentate nucleus.

Statistical Analysis

Demographics and clinical scores were compared between PSP and controls using a Fisher exact test for categoric values and a Wilcoxon rank-sum test for continuous values. They were also compared across PSP clinical variants using a χ² test for categoric values and a Kruskal–Wallis test for continuous values.

The intersubject variability was compared both visually and quantitatively between before and after PC1 subtraction. The intersubject variability was defined as the SD across participants in each group in the 3 composite areas.

The SUVR before and after PC1 subtraction was compared between PSP and controls in the 12 regions. Multiple linear regression analyses were performed in each region, using age and sex as covariates. P values in the multiple regions were adjusted using the false discovery rate (FDR) correction. We repeated the same analysis after adding scanner manufacturer (GE or Siemens) as an additional covariate to evaluate its effect on the comparison. The area under the receiver operating characteristic curve (AUC) was calculated for the 12 regions and 3 composite areas. The DeLong test was used to compare the AUC before and after PC1 subtraction. Voxel-based analyses were also performed to compare SUVR between PSP and controls using statistical parametric mapping (SPM12) to complement the regional analysis. Age and sex were included as covariates in the model. Results were presented at an uncorrected P value of less than 0.001, with a cluster threshold of 50 voxels.

The SUVR was compared between PSP clinical variants (PSP-RS, PSP-subcortical, and PSP-cortical) using a Kruskal–Wallis test. P values were adjusted again using the FDR correction. A post hoc pairwise comparison was conducted using the Dunn test with the Bonferroni adjustment.

The annualized change rate of the SUVR was compared between PSP subgroups (43 PSP-RS, 22 PSP-subcortical, and 9 PSP-cortical) and controls (n = 23) who underwent 2 scans (median scan interval, 1.0 y; interquartile range, 0.99–1.1 y). P values were calculated using the Kruskal–Wallis test with FDR correction and a post hoc Dunn test.

The SUVR was also compared between pathologic diagnoses (PSP, cortical PSP, CBD, and MSA) using the Kruskal–Wallis test with FDR correction and a post hoc Dunn test. Voxel-based analyses were also performed to compare SUVR between CBD (n = 6) and controls. Finally, regional SUVR was correlated with the severity of tau pathology in the corresponding region using Spearman correlation analysis. Note that the supplementary motor area was excluded from this analysis because it did not have corresponding regions in the pathologic analysis. For comparisons involving pathologic diagnoses or tau severity, the SUVR from the second visit was used for the patients who underwent 2 scans.

All statistical analyses were performed using R software version 4.2.2 or MATLAB 2022a.

RESULTS

In the comparison between PSP and controls, the proportion of women was higher in controls than in PSP (Table 1). Education and MoCA were lower in PSP, with no difference in age. In the comparison across PSP clinical variants, there were no differences in sex, education, age at PET scan, or onset. There were differences in disease duration, where the PSP-subcortical variant showed greater duration than the other 2 subtypes. The PSP-cortical variant had more impaired scores on MoCA and Frontal Assessment Battery than the other 2 groups. The PSP saccadic impairment scale also showed differences where PSP-RS and PSP-cortical variants had more impaired scores than did the PSP-subcortical variant. Similar differences were observed in the longitudinal cohort (Supplemental Table 1).

TABLE 1.

Demographics and Clinical Findings for PSP and Control Groups

	PSP and control			PSP clinical variants
Parameter	Control (n = 40)	PSP (n = 141)	P	PSP-RS (n = 80)	Subcortical (n = 44)	Cortical (n = 17)	P
No. of women	29 (73%)	63 (45%)	0.002	34 (44%)	22 (50%)	7 (41%)	0.69
Education (y)	16 (15, 18)	14 (12, 16)	0.01	16 (13, 16)	14 (12, 16)	15 (12, 18)	0.70
Age (y)	69 (66,73)	70 (66, 75)	0.65	70 (65, 75)	71 (67, 75)	70 (65, 76)	0.86
Age at onset (y)		66 (61, 70)		66 (62, 71)	65 (60, 70)	68 (62, 72)	0.47
Disease duration (y)		3.6 (2.4, 5.4)		3.1 (2.2, 4.3)	5.0 (3.3, 7.4)	3.3 (2.3, 6.1)	<0.001
MoCA, n = 30	27 (26,28)	24 (19, 26)	<0.001	23 (19, 25)	26 (22, 27)	20 (18, 26)	0.01
FAB, n =18		14 (12, 16)		14 (12, 15)	15 (13, 17)	12 (9, 15)	0.01
MDS-UPDRS III, n = 132		40 (30, 52)		40 (30, 52)	40 (31, 53)	42 (23, 62)	0.99
PSP rating scale, n = 100		37 (29, 45)		38 (30, 46)	34 (27, 42)	40 (27, 57)	0.07
PSIS, n = 5		3 (2, 4)		3 (2, 4)	2 (1, 3)	3 (2, 3)	<0.001
Scanner manufacturer			0.06				0.54
GE	25 (63%)	111 (79%)		61 (76%)	35 (80%)	15 (88%)
Siemens	15 (38%)	30 (21%)		19 (24%)	9 (20%)	2 (12%)
Pathology							0.004
PSP		23		17	6	0
Cortical PSP		3		0	1	2
CBD		6		2	1	3
MSA		4		0	3	1
Other		2		1	1	0

FAB = Frontal Assessment Battery; MDS-UPDRS = Movement Disorders Society-sponsored revision of the Unified Parkinson Disease Rating Scale Part III; PSIS = PSP saccadic impairment scale.

Qualitative data are number with percentage in parentheses. Continuous data are presented as median with first and third quartiles in parentheses.

PSP Group Comparisons

Figure 2A shows the SUVR images of representative patients with PSP in the template space. Cortical uptake was more homogeneous among patients after PC1 subtraction. Variability decreased after PC1 subtraction by 46% (PSP) and 64% (control) in the frontal GM, 55% and 73% in the frontal WM, and 40% and 56% in the subcortical areas (Fig. 2B). Figure 3 shows representative SUVR images for control, PSP-RS, PSP-subcortical (PSP-parkinsonism), and PSP-cortical (PSP-speech/language) participants in the subject space. All patients with PSP showed uptake in the basal ganglia higher than that in the controls after PC1 subtraction, whereas the patient with PSP-subcortical variant showed an uptake comparable to or smaller than that of the control before PC1 subtraction. Only the PSP-cortical variant showed frontal uptake after PC1 subtraction, whereas the control also showed mild uptake in the frontal before PC1 subtraction.

FIGURE 2.

Intersubject variability. (A) Representative SUVR images in template space in 16 randomly selected patients with PSP. Yellow and green arrows indicate images with relatively higher or lower uptake, respectively. (B) SD across participants in 3 composite areas.

FIGURE 3.

Representative SUVR images for control, PSP-RS, PSP-subcortical (PSP-parkinsonism), and PSP-cortical (PSP-speech/language) participants overlaid on T1-weighted images. Arrows indicate greater uptake.

Figure 4A and Supplemental Table 2 show regional SUVRs in PSP and controls. Before PC1 subtraction, PSP exhibited SUVR higher than that in the controls in 4 subcortical regions: the pallidum, subthalamic nucleus, red nucleus, and cerebellar dentate (FDR-adjusted P < 0.05). After PC1 subtraction, higher SUVR was observed in more cortical and subcortical regions in PSP, including the GM precentral, WM precentral, WM supplementary motor area, WM superior frontal, thalamus, and substantia nigra, in addition to the same 4 subcortical regions. These results remained the same when the scanner manufacturer was added as a covariate. The AUC was improved by PC1 subtraction in the frontal WM (0.54 to 0.67; P = 0.002) and subcortical (0.69 to 0.87; P < 0.001) areas (Fig. 4B).

FIGURE 4.

SUVR comparison between PSP and controls. (A) Boxplots for 12 regions. (B) Receiver operating characteristic curve for 3 composite areas. Sup = superior; Supp = supplementary.

Figure 5 shows voxel-based analysis comparing PSP and controls. Before PC1 subtraction, PSP exhibited higher uptake in subcortical regions, including the basal ganglia, thalamus, midbrain, and cerebellar dentate (P < 0.001). After PC1 subtraction, PSP exhibited higher uptake in the same subcortical regions, with larger clusters and higher t-statistics (voxelwise group comparison statistics), as well as in the frontal areas such as precentral, superior and middle frontal gyri, and the supplementary motor area (P < 0.001).

FIGURE 5.

Regions showing PSP had uptake greater than that in controls (uncorrected P < 0.001). MRIcroGL (version 1.2.20220720) was used for visualization. t = t-statistics.

Figure 6 and Supplemental Table 3 show regional SUVRs across PSP clinical variants. Before PC1 subtraction, 2 subcortical regions showed differences across variants (FDR-adjusted P < 0.05), with greater SUVR in the red nucleus and cerebellar dentate in PSP-RS than in the PSP-subcortical variant. After PC1 subtraction, 3 frontal WM and 5 subcortical regions showed differences across variants (FDR-adjusted P < 0.05). In all 3 frontal WM regions, the PSP-cortical variant showed an SUVR higher than that of PSP-RS and PSP-subcortical variants. In the thalamus, substantia nigra, red nucleus, and cerebellar dentate, PSP-RS showed an SUVR higher than that of the PSP-subcortical variant. In the subthalamic nucleus, PSP-RS showed an SUVR higher than that of both PSP-subcortical and PSP-cortical variants.

FIGURE 6.

Boxplots showing SUVR differences across PSP clinical variants. Sup = superior; Supp = supplementary.

Longitudinal Change over Time

Supplemental Figure 2 shows annualized changes in SUVR. Before PC1 subtraction, there were no significant differences between PSP subgroups and controls. After PC1 subtraction, the subthalamic nucleus showed differences between groups (unadjusted P < 0.05), where the PSP-cortical variant exhibited greater annualized changes than did the PSP-subcortical variant. No regions survived FDR correction.

Associations with Pathology

Figure 7A shows the comparison across pathologic diagnoses. Before PC1 subtraction, no difference was observed after the FDR correction. After PC1 subtraction, differences were observed in the 3 regions (FDR-adjusted P < 0.05). In the WM supplementary motor area, the PSP-cortical variant and CBD showed higher SUVR than did MSA. In the subthalamic nucleus, PSP showed higher SUVR than did MSA. In the red nucleus, PSP showed higher SUVR than did CBD and MSA. In the voxel-based analysis comparing CBD and controls, CBD exhibited greater uptake in the frontal WM and subcortical regions, with larger clusters and higher t-statistics after the PC1 subtraction (Fig. 7B).

FIGURE 7.

Relationship with pathologic diagnoses. (A) Differences in SUVR between pathologic diagnoses. (B) Regions showing CBD had greater uptake than controls (uncorrected P < 0.001). t = t-statistics.

Table 2 shows the correlation coefficients between SUVR and the severity of tau pathology. Before PC1 subtraction, only the red nucleus survived the FDR correction, where greater SUVR was correlated with greater severity of neurofibrillary tangles. After PC1 subtraction, the precentral WM, superior frontal WM, and red nucleus survived the FDR correction. In the precentral WM, greater SUVR was correlated with greater severity of astrocytic lesions. In the frontal WM, greater SUVR was correlated with greater severity of neurofibrillary tangle and astrocytic lesions. In the red nucleus, greater SUVR was correlated with greater severity of neurofibrillary tangle, coiled body, and astrocytic lesions.

TABLE 2.

 Spearman Correlation Coefficients Between SUVR and Tau Pathology Severity

	Before PC1 subtraction				After PC1 subtraction
Parameter	NFT	Threads	Coiled	Astrocyte	NFT	Threads	Coiled	Astrocyte
Precentral (GM)	−0.01	0.30	0.12	0.22	0.03	0.18	0.28	0.39*
Frontal Sup (GM)	0.01	0.24	−0.06	0.05	−0.20	−0.12	−0.18	−0.17
Precentral (WM)	0.17	0.36*	0.16	0.27	0.38*	0.32	0.41*	0.63 †
Frontal Sup (WM)	0.23	0.43*	0.14	0.35*	0.46 †	0.39*	0.39*	0.56 †
Pallidum	0.22	0.33	−0.06	−0.03	0.32	0.36*	−0.09	−0.04
Subthalamic nucleus	0.32	0.13	−0.04	0.12	0.31	0.10	0.26	0.39*
Thalamus	0.12	0.06	0.01	0.10	−0.03	−0.01	0.07	0.11
Substantia nigra	0.34*	0.38*	−0.02	−0.02	0.29	0.39*	0.02	−0.02
Red nucleus	0.54 †	0.36*	0.44*	0.51*	0.48 †	0.34*	0.51 †	0.60 †
Cerebellar dentate	0.25	0.26	0.11	−0.06	0.31	0.26	0.11	0.06

^*Uncorrected P < 0.05.

^†FDR-corrected P < 0.05.

NFT = neurofibrillary tangle; Sup = superior.

DISCUSSION

In this study, we developed a postprocessing method to estimate and subtract the first principal component (i.e., PC1) from [¹⁸F]flortaucipir images and evaluated its effect in a large PSP cohort. The PC1 subtraction greatly reduced (40%–73%) the intersubject variability and improved the ability of [¹⁸F]flortaucipir PET to differentiate PSP and controls, differentiate PSP clinical variants, and differentiate PSP pathology from non-PSP pathology. The PC1 subtraction also improved the correlation between [¹⁸F]flortaucipir uptake and tau burden at autopsy. These results suggest that our subtraction algorithm enhances the diagnostic utility of [¹⁸F]flortaucipir PET in PSP and increases sensitivity to 4R tau, potentially improving the value of [¹⁸F]flortaucipir imaging in non-AD tauopathies.

By PC1 subtraction, the uptake pattern of PSP extended to the frontal cortex, in addition to the subcortical areas. Indeed, frontoparietal regions are affected pathologically in PSP (6,15). The diagnostic performance as measured by the AUC significantly improved in both frontal and subcortical areas. In addition, differences between clinical variants were greatly enhanced by PC1 subtraction. After PC1 subtraction, the PSP-cortical variant had an uptake greater than that of PSP-RS and PSP-subcortical variants in the frontal areas, and PSP-RS had an uptake greater than that of the PSP-subcortical variant in the subcortical regions, which generally aligns with the pathologic tau pattern of clinical variants (6,15). Hence, PC1 subtraction improves the [¹⁸F]flortaucipir signal in PSP.

Longitudinal analysis did not show strong evidence for the longitudinal increase in [¹⁸F]flortaucipir uptake. The results only showed that the PSP-cortical variant had greater annualized changes in the subthalamic nucleus than did the PSP-subcortical variant after PC1 subtraction, but this difference did not survive FDR correction. To our knowledge, only one study from our group investigated longitudinal changes in [¹⁸F]flortaucipir uptake, which showed mild increases over time (16), although its sample size (n = 16) was much smaller than that in our study (n = 74). One possible explanation for our lack of findings may be that tau deposition occurs rapidly in the early stages and then becomes plateaued, as suggested in other 4R tauopathies (17). Another possible explanation could be the partial-volume effects, which may cause cancellation of decreased volumes and increased uptake. Further studies will be needed to clarify if our findings reflect actual tau changes or the limited ability of [¹⁸F]flortaucipir PET to detect subtle increases.

The PC1 subtraction also improved the ability to differentiate pathologic diagnoses and the correlation with the severity of tau pathology. After PC1 subtraction, PSP or PSP-cortical pathology exhibited higher uptake than did MSA in multiple regions, and PSP exhibited higher uptake than did CBD in the red nucleus. These differences align with our understanding of tau profiles in different pathologies (18). The PC1 subtraction enhanced the differences between CBD pathology and controls in the regions expected from the pathologic findings (18). Correlation analysis also showed an improved association with tau severity. These results with pathologic data support the validity of our approach and suggest improved detectability of 4R tau.

We used the cohort from young healthy controls to obtain the PC1 basis image based on the assumption that this component is not related to the diagnosis or age, as suggested by previous studies (8,11,12). This hypothesis is supported by the fact that the obtained basis image worked well in the evaluation cohort, including different diagnoses (PSP) and ages from the young control cohort. In addition, the estimated coefficient for PC1 (k₁) did not associate with age, sex, scanner manufacturer, or diagnosis (PSP/control) in the evaluation cohort (data not shown). Still, it is not fully understood regarding the signal source of PC1, and further research is needed to clarify this. One possible explanation is that individual differences or noise in the reference region may be related to PC1s. Although optimizing the reference region for PSP could address this issue, slight variations in any reference region can affect the entire SUVR signal, potentially leading to increased intersubject variability. The current algorithm offers an alternative and complementary approach to improving SUVR.

The strength of our algorithm lies in its simplicity, unbiasedness, and effectiveness in improving tau PET. Compared with other effective approaches for improving [¹⁸F]flortaucipir PET for AD (19,20), this method does not require additional scans or specific brain atlases. The process was applied in subject space, enabling any further steps such as visual evaluation, voxel-based analysis, and regional analysis. Since the method does not use machine learning or multivariate techniques and relies solely on 1 parameter (k₁), it maintains high levels of explicability and effectiveness. Since the independent cohort does not include patients with PSP, the method could potentially be applied to data from other non-AD patients, such as CBD or frontotemporal lobar degeneration, or even patients with AD at early stages. As this is a postprocessing method, it can also be applied to retrospective data. Another strength is that the method was evaluated using a large PSP cohort, well characterized by experienced neurologists through detailed clinical testing. The cohort also included pathologic data, which are essential for validating our algorithm.

There are several limitations in this study. First, although no notable side effects of PC1 subtraction were observed in this PSP study, if a patient had a global, disease-related uptake pattern (i.e., a similar pattern to the PC1 basis image), such meaningful signals may also be subtracted. Therefore, further refinement may be necessary when applying this method to diseases that could have a strong, global uptake pattern (e.g., advanced AD). Second, there was the difference in sex between the PSP and control groups, though this was accounted for when comparing the 2 groups. Third, the [¹⁸F]flortaucipir uptake would still include off-target signals even after subtraction, which should be interpreted cautiously. Fourth, detailed testing would be necessary to ensure that the algorithm produces consistent results regardless of the scanner type. Fifth, we only evaluated the algorithm for [¹⁸F]flortaucipir, although it could also be potentially applied to other tracers. In particular, it would be interesting to apply this to second-generation tracers with high 4R affinity (21), such as [¹⁸F]PI-2620 and [¹⁸F]APN-1607, to determine whether the algorithm can further enhance sensitivity to 4R tau. Finally, we did not perform a systematic visual assessment of the images. To our knowledge, there are no standardized protocols for visually assessing [¹⁸F]flortaucipir images in PSP. Further studies will be required to establish the visual assessment protocol in PSP, which would enhance the clinical utility of our algorithm in the future.

CONCLUSION

We developed a postprocessing method for [¹⁸F]flortaucipir PET to estimate and subtract PC1 and evaluated its effect in a large PSP cohort. The results indicated that the technique reduced intersubject variability in both cortical and subcortical regions, improving the sensitivity of [¹⁸F]flortaucipir uptake for detecting PSP, differentiating clinical PSP variants, and correlating with the severity of tau pathology. These findings support the effectiveness of our approach and suggest the improved utility of [¹⁸F]flortaucipir imaging in non-AD tauopathies. The proposed method may help improve tau measures as an outcome for potential inclusion in future treatment trials.

DISCLOSURE

This study was funded by NIH grants R01-NS89757, R01-NS112153, and R01-DC12519. Val Lowe consults for Bayer Schering Pharma, Piramal Life Sciences, Life Molecular Imaging, Eisai Inc., AVID Radiopharmaceuticals, and Merck Research and receives research support from GE HealthCare, Siemens Molecular Imaging, AVID Radiopharmaceuticals, and the NIH (NIA, NCI). No other potential conflict of interest relevant to this article was reported.

KEY POINTS

QUESTION: Does [¹⁸F]flortaucipir tau PET detect subtle signals of 4R tau?

PERTINENT FINDINGS: The postprocessing algorithm to subtract the PC1 greatly decreased intersubject variability and improved the sensitivity to 4R tau in terms of separating PSP and controls, separating clinical PSP variants, differentiating pathological diagnoses, and correlating this to tau pathology.

IMPLICATIONS FOR PATIENT CARE: The findings suggest that subtracting the PC1 improved the detectability of subtle 4R tau uptake in PSP and the potential utility of [¹⁸F]flortaucipir PET imaging for non-AD tauopathies.