The necessary of ternary amyloid classification for clinical practice: An alternative to the binary amyloid definition

Visual interpretation is considered the gold standard for amyloid scans in clinical practice. However, dichotomous classification of amyloid deposition by visual reading always results in bias due to rater experience. Therefore, there is a need for a more lenient and flexible amyloid‐equivocal classification in clinical practice. A total of 461 participants were included in this study. Amyloid and glucose positron‐emission tomography was performed, and neuropsychological tests were evaluated. A disease‐specific deep‐learning method was used to identify amyloid equivocality. Amyloid deposition, glucose metabolism, and cognitive performance were analyzed and compared among amyloid‐positive, amyloid‐negative, and amyloid‐equivocal groups. Clinically diagnosed Alzheimer's disease individuals and subjects with normal cognition were used to create amyloid biomarker cut points to support the definition of equivocal amyloid deposition. A total of 139 amyloid‐equivocal individuals were identified by deep learning methods. They displayed intermediate amyloid deposition between that of amyloid‐positive (standardized uptake value ratio [SUVr]: 1.25 ± 0.10 vs. 1.47 ± 0.20, p < 0.001) and amyloid‐negative (SUVr: 1.25 ± 0.10 vs. 1.18 ± 0.07, p < 0.001) individuals. No difference in glucose metabolism or cognitive performance was observed between amyloid negativity and equivocality. Furthermore, the SUVr for the whole cortex, the precuneus, and the frontal lobe served as auxiliary criteria supporting the diagnosis of equivocal amyloid deposition. We also established a guide to assist in the interpretation of amyloid equivocality by visual reading with auxiliary criteria including two cut points and deep learning methods.


INTRODUCTION
Amyloid positron emission tomography (PET) imaging had demonstrated potential applicability in the identification of amyloid-β pathology in patients with Alzheimer's disease (AD) in vivo. 1,2However, amyloid PET image interpretation remains challenging, especially for patients with limited amyloid deposition in specific brain regions.Binary classification has been recommended by many guidelines or associations, such as the Society of Nuclear Medicine and Molecular Imaging and the European Association of Nuclear Medicine. 35][6][7][8] Visual interpretation comprising a qualitative binary reading algorithm that returns a positive or negative result is still considered the gold standard for clinical practice. 3,5isual interpretation is effective for typical positive and negative scans demonstrating striking differences.However, borderline cases with limited amyloid accumulations in very small brain regions may be affected by rater variability. 9For individuals with normal cognition (NC) and patients with mild cognitive impairment (MCI), interpretation of a positive amyloid scan could lead to significant anxiety and even depression.Semiquantitative techniques are often used for research purposes.Many cutoff values have been set for amyloid positivity on amyloid PET imaging, such as a florbetapir standard uptake value ratio (SUVr) (referenced to the cerebellum) greater than 1.10 based on the upper 95% confidence interval of SUVR in a control sample of young healthy participants, as well as an 11 C-PIB SUVr (referenced to the cerebellum) greater than 1.40-1.50based on 90th or 95th percentile of AD distribution. 10n the 2018 National Institute on Aging-Alzheimer's Association (NIA-AA) research framework, a three-range approach was proposed for the biological definition of AD by defining both a lenient and a conservative cut point.11 The three ranges could be labeled as clearly normal (negative), intermediate (equivocal), and clearly abnormal (positive).But some investigations have defined an equivocal amyloid scan by visual reading.Equivocal scans could represent amyloid deposition slightly higher than or similar to that in more than one cerebral gray matter region or a lack of consensus among different raters.6,9 Amyloid-equivocal participants also displayed worse executive performance than amyloid-negative participants.7 these equivocal amyloid depositions were classified subjectively, and the accuracy was largely dependent on the experience of the rater.
Although the cortical SUVr range was also proposed for establishing cut points for amyloid status, this would still be difficult for clinical practice. 5,7In this study, we used a disease-specific deep-learning method to search for intermediate amyloid deposition.This method can implicitly suppress the influence of the non-diagnosis-related normal parts of the brain and thus achieve good performance in AD diagnosis when inputting whole images.We used this method to objectively identify equivocal participants to determine the potential difficulty in diagnosing subclinical participants.After we investigated the status of this group of participants, we proposed a guide to define equivocal amyloid deposition in clinical practice.

Participants
Data used in this study were obtained from the Chinese Preclinical Alzheimer's Study.A total of 461 participants, including 277 NC, 209 MCI, and 75 AD individuals, were recruited from communities in Shanghai, as shown in Figure 1.The exclusion criteria of this cohort included (1) 50-80 years old and provided written informed consent, (2) low education level (≤5 years), (3) presence of neurological or psychiatric antecedents, ( 4) diseases involving the central nervous system, ( 5) severe diseases such as cancer, and (6) significant alcohol and drug abuse.APOE genotyping was performed with polymerase chain reaction and sequencing.This study was approved by the Institutional Ethics Reviewing Board of Huashan Hospital and Shanghai Jiao Tong University Affiliated Sixth People's Hospital.

Neuropsychology
Comprehensive neuropsychological assessments based on the Chinese background were conducted in Shanghai Jiao Tong University Affiliated Sixth People's Hospital by trained neuropsychologists as described in a previous study. 12][15] Diagnoses were made based on clinical examinations according to the corresponding international criteria rather than magnetic resonance imaging (MRI) or PET findings.AD was determined by the 2011 NIA-AA diagnostic criteria for probable AD dementia.MCI was determined by the actuarial neuropsychological method proposed by Jak and Bondi [16][17][18] as two impaired scores (defined as >1 standard deviation below the agecorrected normative mean) in at least one cognitive domain or at least one impaired score in all cognitive domains.The remaining participants were classified as having NC.
F I G U R E 2 Scan reading by a deep learning method: 6-fold cross-validation scheme with a 5-fold partial voting ensemble.

Visual assessment of PET scans
All amyloid-ß-PET scans were read by three experienced, trained radiologist: Donglang Jiang independently who were blinded to the cognitive and clinical diagnosis, as well as clinical information based on recommended practical guidelines. 3The final dichotomous classification (amyloid-ß+ vs. amyloid-ß-) was made by majority consensus.

Scan reading by deep learning method
As described in previous studies, specific brain disease is often associated with certain specific regions, and different regions in the brain vary in anatomy or function.Hence, conventional studies usually partition the brain into multiple regions of interest (ROIs) and then construct a disease classification model using the features extracted in predefined disease-associated ROIs.In contrast, a diseaseimage-specific neural network (DSNN) 21,22 can directly extract features from the whole brain image and find critical information implicitly (e.g., disease-related PET regions) and is thus impacted less by subjective experience.
The DSNN used in this work is derived from our previous work, which has a common structure, as shown at the top right in Figure 2. 21 It assumes that each PET image can be decomposed into (1) a disease-related part and (2) a residual normal part, which is embedded in the feature map extracted by convolutional layers in the DSNN.To make the disease-related features conspicuous and easy to capture, a spatial cosine module with dual-level l 2 -normalization is designed to suppress the effect of the residual normal part.The input of the DSNN is a whole PET image, and the output is a probability score indicating whether this image is positive (1) or negative.While the DSNN is undergoing training, the gold standard label is the visual assessment.
During the application of the DSNN, we designed a 6-fold cross-validation (CV) scheme with a 5-fold partial voting ensemble.An overview of our pipeline is depicted in Figure 2. Our collected samples are first partitioned into exist many inconsistent cases, inapposite criteria of label assignment are implied to be used, for which, we should explore more criteria.

Data preprocessing
The PET images were processed using SPM12.Briefly, PET images were coregistered to the structural MRI, and partial volume error correction was performed using the Muller-Gartner method. 23All images were further warped into the standard Montreal Neurological Institute stereotactic space by using transformation parameters used for warping the MR images.Then, an [8 8 8] full width at half maximum Gaussian kernel was used to smooth the PET data.The mean signals of the bilateral cerebellum crus, whole cerebellum, and brainstem as reference brain regions were used for voxelwise analysis and florbetapir SUVr calculation, while the whole cerebellum was used as the reference brain region for FDG. 24,25

Amyloid cut points for identifying clinically diagnosed AD and NC
First, the cut point corresponding to 90% sensitivity was estimated as the 90th percentile of the AD distribution to search for AD with amyloid-positive deposition, as in a previous report. 10The specificity-based cut point corresponded to the 95th percentile (95% specificity) of amyloid distribution among individuals with normal cognition who had negative amyloid deposition. 26In this method, we only included AD with amyloid deposition and NC without amyloid deposition so that these groups of people were not contaminated by individuals without or with AD pathology.

Statistical analysis
To determine significant differences in demographics and clinical outcomes, the

Demographic information and disease image-specific neural network results
We included 277 NC, 109 MCI, and 75 AD patients in this study (Table S1 and Figure 1).No differences were found in age or sex proportions among NC, MCI, and AD individuals.However, significant differences were observed in education duration and all neuropsychological scores.The AD group displayed a significantly higher proportion of ApoE ε4 carriers than the NC (44.0% vs. 20%, p < 0.001) and MCI groups (44.0% vs. 25%, p = 0.010).
Table 1 shows the demographic characteristics of the subjects divided by amyloid status.No differences were found in age, education years, or percentage of female participants among amyloid positivity, negativity, and equivocality.However, a significantly higher proportion of ApoE ε4 carriers was observed among amyloid-positive individuals than amyloid-negative (52.0% vs. 15.0%,p < 0.001) and amyloid-equivocal individuals (52.0% vs. 22.5%, p < 0.001).In the stratified analysis shown in Tables S2-S4, significantly younger age (63.0 ± 7.9 vs. 67.1 ± 6.5, p = 0.013) and lower ApoE ε4 carrier proportion (12% vs. 43%, p < 0.001) were observed in amyloid-negative individuals than in amyloid-positive individuals in the NC group.In the MCI group, decreased AVLT-LDR and BNT scores and a higher proportion of ApoE ε4 carriers were found

Amyloid deposition and glucose metabolism in subgroups
Voxelwise analysis was performed to identify regions with severe amyloid deposition in amyloid-equivocal participants (Figure 3A-C).We observed amyloid deposition in almost the entire cortex in amyloid equivocality relative to amyloid negativity, with FDR correction, p = 0.005 or p = 0.001.However, amyloid deposition was concentrated on the precuneus, lateral temporal lobe, and frontal lobe with FEW corrections and p = 0.05.
As shown in Table 2 and Figure 3D,E, we found significant global and regional amyloid deposition differences among the amyloid-positive, amyloid-negative, and amyloid-equivocal groups in the overall cohort.Global and regional glucose hypometabolism was observed in amyloid-positive individuals as with amyloid-equivocal and amyloid-negative individuals.In the stratified analysis shown in Figure 3F-K and Supplementary Tables S5-S7, we observed significant global and almost all regional amyloid deposition differences among amyloid-positive, amyloid-negative, and amyloid-equivocal individuals among NC and MCI participants.In AD, amyloid deposition was significantly higher in amyloid-positive participants than in amyloid-negative and amyloidequivocal participants in all regions except the medial temporal lobe.
No global or regional glucose metabolism differences were found among the three groups of participants in the NC and MCI groups.Interestingly, in AD, we observed significant glucose hypometabolism in the lateral parietal lobe, medial temporal lobe, and precuneus in amyloidpositive participants compared with amyloid-negative participants and significant glucose hypometabolism in the posterior cingulate in amyloid-positive participants compared with amyloid-equivocal participants.

Amyloid cut points
Cut points were calculated to support the definition of amyloid equivocality.for the florbetapir SUVr for AD, MCI, and NC participants using bilateral cerebellum crus, whole cerebellum, and brain stem as reference regions.When sensitivity was set to 90% for amyloid positivity and specificity to 95% for amyloid negativity, more amyloid-equivocal, amyloidpositive, and amyloid-negative individuals matched with the artificial intelligence (AI) calculations, as indicated in Table S8.Therefore, SUVr ranging from 1.252 to 1.299 (cerebellum crus as reference), from 1.146 to 1.166 (cerebellum as reference), and from 0.743 to 0.812 (brainstem as reference) could be used to support the definition of amyloid equivocality with different reference regions.For the precuneus (SUVr = 1.303-1.345,lateral temporal lobe (SUVr = 1.254-1.309),and frontal lobe (SUVr = 1.231-1.281),which displayed significant amyloid deposition in amyloid equivocality compared with amyloid negativity, they also need to be noticed.

Clinical guide for amyloid-equivocal scans interpretation
Finally, we established a guide to interpret amyloidequivocal scans in clinical practice, as shown in Figure 5.
First, visual reading is still considered the gold standard.An equivocal amyloid scan can be decided by an experienced senior rater according to the presence of slightly higher amyloid deposition in more than one brain region(s).More commonly, at least two raters had different interpretations leading to a possible decision of equivocal amyloid scan; in these cases, the final decision can be made by a senior rater or consensus.Furthermore, Step 2 is strongly recommended, that is, a review of some specific regions, such as the precuneus and frontal lobe, for a diagnosis of equivocal amyloid deposition.Significant amyloid deposition in more than two brain regions could also be considered amyloid equivocality.At the final step, When MR is available, global and regional amyloid SUVr are auxiliary criteria for supporting the decision.When MR is not available, artificial regional amyloid SUVr is also recommended.An advanced AI method could be applied to confirm the equivocal amyloid deposition diagnosis.

DISCUSSION
Interpretation of amyloid scans is still challenging and relies on the rater's experience when visual reading is  considered the gold standard. 27However, the cut point of global amyloid deposition using SUVr was influenced by preprocessing differences, possible regional heterogeneity of tracer uptake, and choice of the reference region.Therefore, both methods lead to large proportions of false negative/positive interpretations, which may result in inappropriate follow-ups and anxiety for the patient.The misinterpreted cases were always close to the quantitative cut points.Therefore, flexible equivocal amyloid status could provide more options for patient management.We need to pay more attention to individuals with equivocal amyloid deposition, they could have a higher risk to convert to AD compared to NC.
In the 2018 NIA-AA research framework, a two-cutpoint AD biomarker profile was proposed.The AD biomarkers, such as amyloid, might look like A(0 or -), A(1 or equivocal), or A(2 or +).Although some studies have investigated amyloid equivocality by visual reading, this method could produce bias in the definition of amyloid equivocal individuals without validation and further induce different results in research analysis or clinical practice.Therefore, a guide that defines equivocal amyloid is critical for clinical practice.
In this study, we used a disease-specific deep learning AI method to identify equivocal amyloid individuals.This method provided an accurate definition of amyloid equiv-ocality.The advantage of this method is that it can exclude subjective bias by the direct inputting of whole images and identifying samples with incongruous labels (i.e., from visual assessment).With this method, our 5-fold partial voting scheme had difficulty confirming 139 scans.Of these 29 scans resulted in 1 inconsistent result, and 24, 18, 20, and 48 scans generated 2, 3, 4, and 5 inconsistent results, respectively.These 139 scans were defined as equivocal amyloid deposition; the classifications of the remaining scans were consistent with the conventional clinical visual assessments.
After we identified the group with amyloid equivocality, we further investigated the clinical and biomarker properties of this group of participants to confirm the necessity of defining amyloid equivocality.When we divided them by amyloid status, the amyloid-positive group displayed worse neuropsychological scores, global and regional amyloid deposition, and glucose metabolism than the amyloidnegative and amyloid equivocal groups.However, we also found that amyloid deposition in amyloid equivocality was significantly higher than that in amyloid negativity, but no difference was found in neuropsychological scores and glucose metabolism.Amyloid-β production is the earliest pathological process in AD, presenting approximately 15-20 years prior to the clinical symptoms of AD. 28 Glucose metabolism and cognitive performance are regarded as neurodegenerative biomarkers present in the late stages of AD.In this study, we found amyloid deposition but not impaired neurodegeneration in the amyloid equivocal stage.This means that the amyloid equivocal stage could be an intermediate stage between amyloid positivity and negativity.
In the stratified analysis, we found significant global and regional amyloid deposition differences among the amyloid-positive, amyloid-negative, and amyloid-equivocal groups but no difference in cognitive performance or glucose metabolism among them in NC individuals.However, amyloid deposition differences between amyloid negativity and equivocality were only found in the frontal, temporal lobe, and precuneus and no glucose metabolism difference was found between these two groups in MCI.In AD, no amyloid deposition differences were found, but a glucose difference was observed in the posterior cingulate between amyloid equivocality and negativity.We found amyloid equivocality only displayed amyloid deposition in preclinical stages of AD, neurodegeneration emerges in AD.It demonstrated early detection of amyloid deposition has the potential for precise management of this disease.And the definition of amyloid equivocality is critical.
To set up a workflow for the definition of equivocal amyloid in clinical practice, we further sought to identify some auxiliary criteria.The first auxiliary criterion could be the definition of two cut points as proposed by the 2018 NIA-AA research framework.We established the two cut points auxiliary criteria using cerebellum crus, whole cerebellum, and brain stem as reference regions.However, these cut points were conservative, and a lenient and flexible range is also acceptable.In clinical practice, MR is not always available, assessment of amyloid uptake from CT is also helpful.However, in this situation, whole cortical amyloid uptake calculation is not possible.Voxelwise analysis revealed that the precuneus, frontal lobe, and lateral temporal lobe displayed higher amyloid deposition in a strict analysis (p = 0.05, FWE corrected).Additional attention should be given to these regions to distinguish amyloid equivocality from amyloid positivity and negativity in clinical practice.Therefore, cut points for these specific brain regions are necessary.Previous results also indicated that the precuneus is the earliest cortical region for amyloid accumulation, followed by the frontal lobe and lateral parietal and temporal lobes in AD. 29 In conclusion, cut points for these regions are also important for supporting a diagnosis of amyloid equivocality.
The equivocal amyloid deposition could be an intermediate stage between positive and negative amyloid deposition stages, in the same way, that MCI is an intermediate stage between NC and AD.An amyloid-equivocal diagnosis could be a risk biomarker for amyloid deposition.A diagnosis of equivocal amyloid deposition could reduce panic more than an amyloid-positive diagnosis.There were several limitations in this study.First, we did not have follow-up data to confirm the risk of the amyloid equivocal groups.It is necessary to check the conversion status of these individuals, as well as their cognitive performance and biomarker alterations.In the previous study, approximately 8% of visually equivocal amyloid PET scans with intermediate load had a rapid accumulation of amyloid during the follow-up. 30Therefore, the longitudinal results of equivocal amyloid in a larger cohort were critical.Second, the sample size was not large enough, especially for AD patients.Because all of these volunteers were from communities, the incidence of AD is not as high as that in memory clinics.A larger sample size will be helpful for future studies on this matter.

CONCLUSION
In his study, we used a deep learning method to identify a group of individuals with equivocal amyloid deposition.These individuals displayed intermediate amyloid deposition between amyloid positivity and negativity, but no difference was observed in cognitive performance and glucose metabolism among amyloid equivocality, negativity, and positivity.Finally, we established a guide for identifying amyloid equivocality in clinical practice by visual reading with auxiliary criteria to support the decision.

A U T H O R C O N T R I B U T I O N S
Fang Xie, Qihao Guo, and Dinggang Shen designed this study and coordinated all the research.Fang Xie and Qihao Guo organized the data collection; Shuhua Ren, Junpeng Li, Lin Huang, Qi Huang, and Donglang Jiang collected the data; Shuhua Ren, Yongsheng Pan, Junpeng Li, and Qi Huang processed and analyzed the data; Yihui Guan, Qihao Guo, and Fang Xie reviewed the study design and redesigned this study.Fang Xie led the manuscript writing, and all authors reviewed and revised the manuscript.

A C K N O W L E D G M E N T S
The authors want to thank Jianfei Xiao, Yunhao Yang for the generous assistance in tracer production and Xiangqing Xie, Zhiwei Pan, Yue Qian, and Dan Zhou for their assistance in the patient arrangement.

Figure 4 illustrates individual values F I G U R E 3
Amyloid deposition and glucose metabolism in amyloid equivocality versus amyloid positivity and negativity.(A-C) Voxelwise analysis for amyloid deposition in amyloid equivocality versus amyloid negativity in different conditions; (D, E) Amyloid deposition and glucose metabolism for all individuals; (F-K) Amyloid deposition and glucose metabolism for normal cognition (NC), mild cognitive impairment (MCI), and Alzheimer's disease (AD).

TA B L E 2
Amyloid deposition and fludeoxyglucose (FDG) accumulation in different brain regions in all participants.Presented as the mean (SD).

F I G U R E 4
Cut points for three-range classification.Arrows in each panel indicate the biomarker cutoff points for 2 levels of Alzheimer's disease (AD) sensitivity and normal cognition (NC) specificity ranging from 90% to 95%.(A-C) Whole cortical SUVr cut points with reference to the cerebellum crus, cerebellum, and brain stem.(D-F) Brain regional cut points for the precuneus, frontal lobe, and lateral temporal lobe using the cerebellum crus as a reference.

F I G U R E 5
Guide for defining equivocal amyloid deposition in clinical practice.
This research was sponsored by the National Key R&D Program of China (2016YFC1306305, 2018YFE0203600), STI2030-Major Projects (2022ZD0213800), the National Science Foundation of China (81801752, 82171473, 82201583, and 82071962), the Shanghai Sailing Program (18YF1403200 and 19YF1405300), the startup fund of Huashan Hospital, Fudan University (2017QD081), the Shanghai Municipal Key Clinical Specialty (3030247006), Clinical Research Plan of SHDC (No. SHDC2020CR2056B), and the Shanghai Municipal Science and Technology Major Project (No. 2018SHZDZX01) and ZJLab.C O N F L I C T S O F I N T E R E S T S TAT E M E N T The authors declare no conflict of interest.O R C I D Fang Xie https://orcid.org/0000-0003-2667-281XR E F E R E N C E S Demographics and clinical outcomes of all participants.Presented as the mean (SD).
TA B L E 1a Significant difference between amyloid + and amyloid -. b Significant difference between amyloid-and amyloid equivocal.c Significant difference between amyloid + and amyloid equivocal.*Significant difference between two groups, p < 0.05.**Significant difference between two groups, p < 0.01.***Significant difference between two groups, p < 0.001. in amyloid positivity than in amyloid negativity.Worse AVLT-LDR performance was also observed in amyloid positivity than in amyloid equivocality in the MCI group.However, no differences in age, sex proportions, education years, neuropsychological scores, or ApoE ε4 carriers were observed among AD individuals.