Radiomics analysis of magnetic resonance imaging helps to identify preclinical Alzheimer’s disease: an exploratory study

Background: Diagnosing Alzheimer’s disease (AD) in the preclinical stage offers opportunities to early intervention, however, there is a lack of convenient biomarkers currently. By using the methods of radiomics analysis, we aimed to determine whether the features extracted from multi-parameter magnetic resonance imaging (MRI) can be used as potential biomarkers. Methods: This study is part of the SILCODE project (NCT03370744). All participants were cognitively healthy at baseline. The cohort 1 (n=183) was divided into individuals with preclinical AD (n=78) and controls (n=105) by amyloid-positron emission tomography, used as the training dataset (80%) and validation dataset (the rest 20%); cohort 2 (n=51) was divided into “converters” and “non-converters” by individuals’ future cognitive status, used as a separate test dataset; cohort 3 included 37 “converters” (13 from ADNI), was used as another test set for independent longitudinal researches. We extracted radiomics features from multi-parameter MRI of each participant, used t-tests, autocorrelation tests, and three independent selection algorithms, respectively, to select features. Then we established two classication models (support vector machine (SVM) and random forest (RF)) to verify the eciency of retained features. Five-fold cross-validation and 100 repetitions were carried out for the above process. Furthermore, the acquired stable high-frequency features were tested in cohort 3 by paired two-sample t-tests and survival analyses, in order to identify whether their levels change with cognitive decline and impact conversion time. Results: The SVM and RF models both showed excellent classication eciency, with the average accuracy of 89.65%-95.90% and 87.07%-90.81% respectively in the validation set, 81.86%-89.10% and 83.19%-83.68% respectively in the test set. Three stable high-frequency features were identied, namely Large zone high-gray-level emphasis feature of right posterior cingulate gyrus, Variance feature of left superior parietal gyrus and Coarseness feature of left posterior cingulate gyrus, each cognitive decline of participants in the test set, the areas under (AUCs) The 13.0 under curve; MoCA: Montreal cognitive assessment; LZHGE: Large zone high-gray-level emphasis. and Cog-M groups. At baseline timepoint, all participants were cognitively healthy, and we made comparisons of clinical data between the two groups of cohort 1 and 2, respectively. The cohort 1 was used as the training and validation dataset, the cohort 2 was used as the test set. The cohort 3 was applied to the longitudinal study, including the 24 participants from cohort 2 (Cog-D) and 13 from ADNI, the latter was a supplement and identical to the evaluation of Cog-D group. The MoCA scale applied in cohort 1 was Chinese MoCA-Basic version, in cohort 2 was MoCA-Beijing version, and in ADNI was traditional MoCA version. Continuous measures were presented conducted Chi-square test for categorical variables independent two-sample two-tailed t

The study was part of the Sino Longitudinal Study on Cognitive Decline (SILCODE), an ongoing prospective cohort study (ClinicalTrials.gov identi er: NCT03370744; protocol can be accessed at ClinicalTrials.gov) [30] , which centers on Xuanwu Hospital, and the alliance includes 94 hospitals from 50 cities in China. The SILCODE project is a constellation of interconnected sub-studies, and one of its aims is to assess the diagnostic application of imaging in different stages of cognitive continuum. Therefore, baseline standardized clinical evaluation and MPMRI were offered to all participants, resulting in the enrollment of 1594 with different diagnoses, ranging from cognitively unimpaired to dementia. In this study, we established three cohorts from the database and Alzheimer's Disease Neuroimaging Initiative (ADNI, www.loni.ucla.edu/ADNI) with high selectivity. In cohort 1, one hundred and eighty-three cognitively healthy participants with amyloid-PET imaging were being included between July 2016 and November 2018 sequentially, they were all from SILCODE project (Supplementary Figure 1A). In cohort 2, fty-one participants were included, they participated in SILCODE project from December 2009 to December 2015 and were selected retrospectively; they were interviewed every 10-15 months until the end of 2019, with 24 had future cognitive decline and 27 remained healthy (Supplementary Figure 1B). In cohort 3, the 24 "converters" from cohort 2 and additional 13 individuals from ADNI were included, they all underwent MPMRI examinations at baseline timepoint and when they were rst identi ed cognitive deterioration. The entry criteria of healthy individuals referred to our previous references [30,31] ; the diagnosis of dementia referred to the guidelines of NIA-AA workgroups [32] ; MCI was based on the Petersen's criteria (before 2016) [33] or a neuropsychological method (after 2016) [34] .
Participants in cohort 1 all underwent a dynamic scan with Florbetapir F-18 (AV45), the whole brain voxel-wise standardized uptake value ratio (SUVR) was normalized to the whole cerebellum, representing the mean cortical SUVR. For the dichotomy, amyloid-PET positivity, that is participants who in the preclinical stage of Alzheimer's continuum, was de ned a priori with the established cutoffs of >1.18 [35] . The result of each participant was con rmed by two senior radiologists.
Informed consent was obtained from all participants. Further details regarding the rigorous evaluation of our participants were presented in the Supplementary Figure 1 and Material.

Imaging acquisition and preprocessing
The MRI data of participants from SILCODE project were acquired using a 3.0-T MR imager (Magnetom Sonata; Siemens, Germany) before the year of 2016, or an integrated simultaneous 3.0-T TOF PET/MR (SIGNA; GE, USA) after that timepoint. Before undergoing imaging, subjects were instructed to keep their eyes closed but not fall asleep, to relax their minds, and to move as little as possible during imaging, and foam pads and headphones were used to minimize head movement and imager noise. The sMRI was obtained with a magnetization prepared rapid gradient echo sequence (Siemens/GE): repetition time (TR) = 1900 ms/6.9 ms, echo time (TE) = 2.2 ms/2.98 ms, slice number = 176/192; fMRI was obtained with a single-shot gradient-echo echo planar imaging (EPI) sequence (Siemens/GE): TR = 2000 ms/2000 ms, TE = 40 ms/30 ms, slice number = 28/28; and a single-shot spin-echo diffusion-weighted EPI sequence was used for the DTI data (Siemens/GE): TR = 11000 ms/16500 ms, TE = 98 ms/95.6 ms, slice number = 60/75. The detailed protocols could be found in the Supplementary Material.
The images of the ADNI participants were downloaded from the ADNI database, detailed information regarding the acquisition protocol is publicly available on the LONI website (https://ida.loni.usc.edu/login.jsp).
The standardized preprocessing of amyloid-PET and MPMRI referred to previous studies [30,36,37] . The original DICOM data were converted to NIfTI le format by using DCM2NII (https://people.cas.sc.edu/rorden/mricron/dcm2nii.html), then we processed the MPMRI and amyloid-PET separately for each participant.
For sMRI, the images were normalized and showed a spatial resolution of 91×109×91 with 2×2×2 mm 3 voxel size after being segmented into gray matter, white matter and cerebrospinal uid tissues, afterwards, we smoothed them by using an isotropic Gaussian smoothing kernel with the full-width at half maximum of 4×4×4 mm 3 . For fMRI, the rst 10 time-point images were discarded for magnetization balance, after that, the remaining 230 time-point images were corrected and aligned to the rst time-point image to correct head movements; the resultant motion-corrected volumes were co-registered to the anatomical T1-weighted images and normalized to the Montreal Neurological Institute template, resampling to a 3 mm cube voxel resolution. For DTI, we employed the Eddy Correct tool to correct the head motion and eddy current distortions [38,39] , and used the Brain Extraction tool to remove the non-brain tissues of the B0 image and create the brain mask [40] , then we adopted the DTIFIT tool to t the diffusion tensor at each voxel and produced four parameter maps, encompassing fractional anisotropy, mean diffusivity, axial diffusivity and radial diffusivity [41] . For amyloid-PET, the structural images were individually registered to the averaged PET images, then we performed segmentation of all the co-registered structural images, spatially normalized the PET images to the Montreal Neurological Institute standard space by using the forward parameters (estimated during the segmentation), smoothed the images with an 8 mm full width at half maximum Gaussian kernel.

Features extraction
The features extraction was performed for each modality separately. For sMRI, 43 texture features and 172 wavelet features of each region of interest (ROI) (116 in total, based on the AAL template) were extracted; for fMRI, 43 texture features of each ROI were extracted; for DTI, we calculated the white matter tracts and viewed the fractional anisotropy, mean diffusivity, axial diffusivity and radial diffusivity as features. All these extracted features were adjusted by using linear regression to control the impacts of age, gender and education before selection.
Features extraction of sMRI and fMRI was performed on the Texture Toolbox from radiomics tools developed by Vallieres et al.

Features dimensionality reduction and selection
This step was achieved on MATLAB. More speci cally, we performed a 5-fold cross-validation on the dataset of cohort 1, that is, the data was randomly divided into a training set (80%) and a validation set (the rest 20%); in the training set, three steps including t-tests, autocorrelation tests and independent three different algorithms (Fisher score, Least absolute shrinkage and selection operator (Lasso) and Max-Relevance and Min-Redundancy (mRMR)) were adopted to lter the redundant and meaningless features; the remaining features were retained and incorporated into classi cation models. Importantly, we repeated the above steps 100 times. More details could be found in the Supplementary Material.
In addition, for each type of above three algorithms, we calculated the number of occurrences of each retained feature, ranging from 0 to 500. The top ten most frequently appeared features were de ned as high-frequency features, and the stable high-frequency features means the overlaps among the three perturbations.

Classi cation experiments
We established two classi cation models, support vector machine (SVM) and random forest (RF), to verify the performance of retained features in the validation (20% of cohort 1) and test set (cohort 2) respectively. The SVM model employed three different kernels, sigmoid, linear and radial basis. Corresponding to the retained features, there were 500 permutation experiments under the algorithm of Fisher score, Lasso or mRMR, then the nal results of accuracy, sensitivity and speci city were presented as average values ± standard deviation (SD) of each model.
In addition, the receiver operating characteristic (ROC) analyses were performed to evaluate the ability of each stable high-frequency feature in predicting prospective cognitive decline of participants in the test set, and the areas under curve (AUCs) were calculated. The analysis was achieved by SPSS 13.0 software (Chicago, IL, USA).
Details could be found in the Supplementary Material.

Longitudinal analyses
As an independent longitudinal research, this study aimed to further verify the role of stable high-frequency features identi ed from the training dataset (80% of cohort 1) in another separate test dataset, that is the cohort 3. The features extraction was identical to above-mentioned. In order to test whether the levels of stable high-frequency features change with cognitive decline, we made comparisons at the two different timepoints of cognitive stages, and in order to verify whether these features have impacts on the conversion times of individuals, we performed the survival analyses.

Statistical analysis
The demographic data of participants were summarized as numbers (%) or mean ± SD for categorical and continuous variables, respectively. The betweengroup comparisons were performed by using chi-square test for categorical variables, or the two-sample t-test for continuous variables (two-tailed). A p<0.05 was considered signi cant.
In the process of dimensionality reduction, the two-sample t-test was two-tailed and considered signi cant when p<0.05; for the autocorrelation test, we calculated the Pearson correlation coe cients between features and considered the paired features had a high correlation when values in the pair-wise correlation were greater than 0.8. Furthermore, in order to better understand the association between radiomics features and iconic pathological changes of AD, we made Pearson correlations to evaluate the relationship between stable high-frequency features and mean cortical SUVR values, and acquired the results after adjusting for age, gender, education, and Montreal cognitive assessment (MoCA) scores.
In the longitudinal analyses, we drew the changing trajectory of each stable high-frequency feature at the individual level, and performed the paired twosample t-tests at the group level (two-tailed, p<0.05). In the survival analyses, individuals of cohort 3 were equally divided into two parts, the high-level group (n = 18) and low-level group (n = 19), according to the median level of each stable high-frequency feature, respectively, afterwards, cumulative probabilities of clinical conversion by the two groups were displayed according to the Kaplan-Meier method, then the survival curves were compared between groups in an univariate analysis applying the log-rank test.
These above analyses were performed in SPSS or MATLAB.

Results
In cohort 1, 183 healthy participants were included. Their clinical and MPMRI examinations were almost continuous, and the amyloid-PET was performed within 3 months of the MPMRI scan. Eventually, seventy-eight amyloid-positive and 105 negative participants were identi ed. Compared to the negative individuals, individuals who were positive were older (p = 0.039) and had higher AV45 SUVR (p<0.0001), while there were no statistical differences of other clinical data (Table 1).
In cohort 2, additional 51 healthy participants were dichotomized due to their future cognitive outcomes. They were interviewed every 10-15 months and we found the cognition of 24 deteriorated after an average of 41.2 months (IQR: 24.5-52.7), with 23 progressed to MCI and 1 to dementia, while the remaining remained healthy after at least 3 follow-up visits (54.8 months, IQR: 48.9-58.9). As shown in Table 1, there were no differences between the two groups. Cohort 3 included the 24 "converters" from cohort 2 and 13 from ADNI, their average score of MoCA scale slumped from 23.7±2.7 at baseline timepoint to 20.5±3.8 at follow-up timepoint. The average conversion time of ADNI participants was 62.1 months (IQR: 55.1-66.5), of total individuals was 48.1 months (IQR: 27.8-61.7). The individuals from ADNI were used as an additional supplement, with 12 progressed from cognitively healthy to MCI and 1 to dementia.
Other data could be acquired in Table 1.

Features extraction, dimensionality reduction and selection
For each participant of the three cohorts, 30128 features were extracted, including 24940 features from sMRI, 4988 from fMRI and 200 from DTI. To avoid over-tting, these features were screened before being included into classi cation models. In the training set, 9,000-11,000 features were retained after the two-sample t-tests (p<0.05) and 2200-2500 types of uncorrelated features were reserved after the autocorrelation tests. The remaining features were further ltered by three independent selection algorithms respectively, more speci cally, we retained the top 50 ranked features by using the Fisher score test, 50-70 features using the Lasso method, and top 50 ranked features after the mRMR test.
Generally, the retained features showed consistency in repeated experiments. As shown in Table 2, there were 10 high-frequency features of each composite function disturbance, notably, they were all based on sMRI modality. For the features selected from the disturbance containing Fisher score test, the frequency was 420-500 times, mainly originated from the posterior cingulate (left, 3/10; right, 3/10); for the features selected from Lasso, the frequency was 383-468 and no speci c regions were identi ed; for the features selected from mRMR, the frequency was 320-495, also mainly from the posterior cingulate (left, 4/10; right, 2/10).
Three stable high-frequency features were identi ed during the process, they were undisturbed by the combined disturbances, and may of great importance in Other retained features that occurred more than 300 times and the meanings of stable high-frequency features were described in the Supplementary Table 2 and Material.

Classi cation experiments
We introduced two types of models to determine whether the retained features were compatible for classi cation analysis. The Table 3 presented the classi er performance results in terms of accuracy, sensibility and speci city. As shown, the SVM model (radial basis kernel) showed excellent classi cation e ciency, with the average accuracy up to 90.23%-95.90% (sensibility, 85.91%-92.82%; speci city, 93.71%-98.26%) in the validation set, and 84.48%-88.94% (sensibility, 79.76%-82.87%; speci city, 85.98%-96.70%) in the test set. Similar results were acquired in the RF model (Table 3) or SVM models with the other two kernels (Supplementary Table 3). In contrast, the average accuracy of pure clinical data-based models in diagnosing preclinical AD reached only random level accuracy, was 55.93%-56.01% (details were presented in the Supplementary Table 4 and Material).
We further veri ed the classi cation e ciency of stable high-frequency features on the test set, and found their individual AUCs ranged from 0.649 to 0.761, when we combined them together, the predictive ability was improved (AUCs = 0.839) (Figure 2A-D). In addition, the feature 6486 also played a good classi cation effect (AUCs = 0.739), and improved the AUCs to 0.863 when we combined it with the three stable features (Supplementary Figure 2). In contract, the performance of feature 28977 was too bad to draw a ROC curve. These results indicated that radiomics analysis was a reliable feature extraction method in the preclinical stage of AD, and provided promising imaging biomarkers in identifying cognitively healthy individuals with future cognitive decline.

Correlation analysis
In order to further understand the association between radiomics features and pathological changes of AD, we performed the correlation analysis between stable high-frequency features and mean cortical SUVR values of amyloid-PET and found they were highly correlated. In detail, the feature 6489 levels were positively correlated with SUVR values (r = 0.433, p<0.0001, Figure 2E), while the feature 11517 and 27442 levels were both inversely correlated with SUVR values (r = -0.416, p<0.0001, Figure 2F; r = -0.348, p<0.0001, Figure 2G). Similar results were found in the feature 6486 (r = -0.400, p<0.0001) (Supplementary Figure 2). The correlation results did not change after adjusting for age, gender, education, and MoCA score (Supplementary Figure 3). Our ndings revealed that high correlations between levels of these features and Aβ depositions, suggesting that radiomics features based on MPMRI may re ect pathological changes in the brain and be used for the diagnosis of AD.

longitudinal analyses
In this study, the 37 participants of cohort 3 were followed up until cognitive impairment was found. First, we detected their longitudinal changes of each stable high-frequency feature. As shown, the feature 6489 and 11517 did not show isotropic changes of the two cognitive stages at the individual level ( Figure   3A, B), correspondingly, there were also no statistical differences between the two paired groups ( Figure 3D, E), similar results were obtained of the feature 6486 (Supplementary Figure 4A, B); although some individuals had a heterogeneous change pattern of the feature 27442 ( Figure 3C), but its levels in the cognitive impairment stage was still lower than that in the cognitively healthy stage (p = 0.0403) ( Figure 3F). Second, we made survival analysis of these features respectively. In detail, the median baseline level of feature 6489, 11517 and 27442 was 0.0297356, 17228.308, 0.865647 respectively; Figure 3G-I showed the probability of cognitive impairment by levels of features > and ≤ these cutoffs, notably, in the comparison between paired groups, only grouping by feature 27442 was meaningful (log rank p = 0.015). The result of feature 6486 was also unsatisfactory when grouped by the median level of 48.967 (log rank p = 0.442) (Supplementary Figure 4C). These results indicated that the levels of feature 27442 decreased with cognitive decline, and the deterioration occurred earlier when the baseline level was less than 0.865647, however, considering the limited sample size, the value is for reference only, and it is more accurate to state that the baseline level can affect the conversion time.

Discussion
The real pathophysiological process of AD is thought to begin several decades before symptom onset, generally followed a rigid progress pattern that is Aβ accumulation-neuro brillary tangles-neuronal damage; neurons had already damaged to some extent when cognitive impairment occurred [1,8] . Radiomics analysis can extract high-dimensional features of MPMRI, may identify imaging patterns of preclinical stage that cannot be recognized by human readers, however, there is a paucity of published literatures assessing radiomics features of individuals in preclinical stage of AD or with future cognitive decline. In this ongoing prospective cohort study, we adopted a novel composite method to select features from the training dataset, established classi cation models and veri ed them in the validation set and test set respectively; we found both models can distinguish whether individuals were in the preclinical stage or whether their future cognition will decline, with the accuracy more than 80%. In addition, three stable high-frequency features were identi ed, they were independent of perturbations and correlated with Aβ deposition, all can classify the test set accurately (AUCs 64.9% to 76.1%). In the independent longitudinal analyses, we further veri ed that levels of the feature 27442 (Variance feature of left superior parietal gyrus on sMRI) decreased with cognitive decline and impacted individuals' conversion time. Together, these data showed that radiomics features of MPMRI were expected to be important imaging biomarkers in accessing patients with preclinical AD.
Our previous researches con rmed that cognitively normal high-risk individuals of AD have already appeared alterations of brain functional networks (fMRI), white matter networks (DTI) or some re ned areas (sMRI) [46][47][48][49] , suggesting there may be more unmined data of MPMRI in the preclinical stage. As expected, the pure clinical data-based classi cation models were meaningless in this stage, and the traditional volumetric and functional indices are also not sensitive enough (details were presented in the Supplementary Table 5 and Material). Although it is generally believed that radiomics analysis is more sensitive, but current studies are still limited to symptomatic stages [18,19,22,23,29] . Chaddad et al. found the features derived from a single subcortical region produced an AUCs up to 80% for classifying AD-dementia from healthy individuals, and reached 91.54% when combined all regions [22] . By using hippocampal features, researchers can distinguish AD-dementia with an accuracy of 86.75%, and 70.51% of MCI [19] . Identical conclusions were obtained in a recent large-scale multicenter study where the hippocampal features served as robust biomarkers for clinical identi cation of AD-dementia/MCI and further predicted whether MCI patients would convert to dementia [23] . By contrast, the deep learning methods indeed can acquire a slightly better diagnostic capabilities in Alzheimer's continuum [50] , however, it is di cult to explain the clinical correlations between these deep features and AD itself, and notably, Li et al. have proved that the performance in identifying dementia from controls by using radiomics was comparable to deep learning (91.4% and 93.9%, respectively) [51] . Here, in distinguishing preclinical AD patients or clinical "converters", the accuracy of our models reached an amazing 81.86%-95.90%, even higher than that of distinguishing symptomatic patients from controls. We thought several reasons may account for it. First, compared with extracting features solely on sMRI, we were based on MPMRI. Second, instead of selecting regions on prior knowledge, we adopted template segmentation and extracted features respectively.
Third, among the huge 30128 features, we used an innovative selection method to improve the robustness. Four, we diagnosed individuals based on the Aβ pro le, not purely on clinical data, signi cantly reduced the heterogeneity of participants. Moreover, different types of models further veri ed the reliability.
The Aβ deposition associated with neuronal degeneration may have resulted in subtle alterations in MRI signal intensity, therefore, we speculate that radiomics features can re ect changes at the microscopic level of early pathological status, which occurred before changes at the macroscopic level. In addition to the computer-aided classi cation, three stable high-frequency features, which were not affected by function perturbations (three different algorithms) and sample perturbations (5-fold cross validation and 100 repetitions), were identi ed during the selection process, they were the LZHGE feature of right posterior cingulate gyrus, Variance feature of left superior parietal gyrus and Coarseness feature of left posterior cingulate gyrus, all on sMRI modality. Importantly, the earliest accumulation of Aβ deposition is also in the superior parietal gyrus and posterior cingulate [1] . More speci cally, in symptomatic AD patients, previous ndings of autopsy and amyloid-PET suggested that parietal lobe and posterior cingulum are vulnerable to Aβ invasion during early stage [52,53] ; in cognitively normal individuals, the annual increase of Aβ also localizes to these two regions [54] . From other perspectives, the Aβ deposition is particularly associated with the cortical atrophy of superior parietal gyrus [55,56] , and the rate-limiting enzyme of Aβ production is also signi cantly elevated in this area [57] . These developments prove the accuracy of locations, support our results that these features were correlated with SUVR values, played good roles in predicting future cognitive decline (AUCs 64.9%, 72.9% and 76.1% respectively; 83.9% when combined), and probably being the imaging biomarkers of preclinical AD. Interestingly, we found the retained features only came from sMRI modality, this is probably in part due to the relatively small number of fMRI and DTI features, additionally, a latest study concluded that DTI parameters is not useful for the identi cation of preclinical AD [58] , and to the best of our knowledge, it is the rst time that texture analysis of fMRI have been applied to the eld of AD [36] . Uncertainty still exists, the signi cance of DTI and fMRI radiomics features cannot be completely denied in the exploratory study.
In the longitudinal analyses, we found levels of the Variance feature of left superior parietal gyrus on sMRI decreased with the impairment of cognition, suggesting it may of great importance in the whole cognitive continuum, not just in the preclinical stage. This feature is extracted from the gray-level cooccurrence matrix category, is an indicator of dispersion of the unit values around the mean. With cognition decline, the cortical accumulation of Aβ will increase continuously in a certain extent [1] , may result in alterations in signal intensity, and then this subtle changes were captured by the radiomics analysis of sMRI. Next, we conducted survival analyses to compare the conversion time between groups within the cohort 3. The median value was chosen subjectively for grouping; coincidently, we found also only the Variance feature can affect the conversion time, further suggested its predictive effects on clinical outcome.
Our study had some limitations. First, the small sample number limited the statistical power of our data. We tried to overcome this issue by enrolled participants from other sub-centers and ADNI, but the requirement of amyloid-PET or long-term follow-up or MPMRI greatly limited the quantity. Moreover, the performance of our models maybe different by using different imaging protocols. Second, considering no standard de nition of "unstable preclinical AD", we referred to the 36 months of "unstable MCI", and required the "non-converters" must remain cognitive stability at least three follow-up visits, while the average conversion time of "converters" was 41.2 months, which needs to be veri ed. Third, other regions, such as anterior cingulate, are also susceptible to Aβ attack [52,53] , however, we did not nd any stable features situated in these regions. Fourth, in cohort 3, most of the patients were limited at MCI stage and few of dementia stage at the follow-up timepoint, thus it is not clear whether features were related to the degree of cognitive deterioration. Fifth, the positive result of study 2 were not that signi cant (p = 0.0403), and the feature levels of some individuals were increased disparately, which probably due to the heterogeneity of MCI and the relatively older age of ADNI participants. Sixth, age may cause differences of our results because of its impacts on Aβ and atrophy. Seventh, the positive rate of amyloid-PET (42.6%) was higher than that of previous researches (10%-30% mostly) [59] , partly because of the exclusion of some negative individuals (Supplementary Figure 1A) and the existence of individuals with subjective cognitive decline, a high-risk state of developing AD [60] ; this bias may increase uncertainty. Considering the shortcomings of our research and limitations in this eld, multicenter collaboration to include more participants are needed in the further.

Conclusions
In conclusion, radiomics analysis of MPMRI is expected to become a new evaluation method for Aβ deposition and future cognitive decline in cognitively healthy individuals, of great importance in diagnosing preclinical AD and targeting super-early secondary prevention clinical trials. Additionally, we proposed a novel paradigm of features extraction and preservation method of features subset, solving the problem of instability and non-repeatability for future studies.

Ethics approval and consent to participate
The study was approved by the Medical Ethics Committee of Xuanwu Hospital of Capital Medical University and was carried out in accordance with the Declaration of Helsinki. We con rm that we have read the Journal's position on issues involved in ethical publication and a rm that this report is consistent with those guidelines.

Consent for publication
All subjects gave written informed consents and written consent to permit publication of clinical details.

Availability of data and materials
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Competing interests
On behalf of all authors, the corresponding author con rms no con ict of interest.

Authors' Contributions
Ying H provided the data; JH Jiang designed the study; TR Li and Y Wu assembled and analyzed the data, consulted literatures and drafted the manuscript; JJ Jiang drawn the Figure 1; Ying H and JH Jiang critically revised the manuscript for important intellectual content. All authors read and approved the nal manuscript.   Under the sample disturbance of 5-fold cross-validation, we have carried on three different kinds of composite function disturbances separately to screen features in the training dataset, and repeated the process 100 times. We calculated the number of occurrences of each retained feature, ranging from 0 to 500, and listed the top ten most frequently appeared features here, they were all originated from sMRI modality. Three stable high-frequency features were veri ed, and their identify numbers were 11517, 27442 and 6489 respectively. The Kurtosis feature belong to the "global" category; the Homogeneity and Variance features belong to the "gray-level co-occurrence matrix" category; the GLN, ZSN, LZHGE, SZLGE, LZLGE and ZSV features belong to the "gray-level size zone matrix" category; the Strength, Coarseness, Busyness, Complexity and Contrast features belong to the "neighborhood gray-tone difference matrix" category.
Notably, the Variance and Contrast feature can also originate from the "global" and "gray-level co-occurrence matrix" category, respectively.  Abbreviations: SVM, support vector machine; RF, random forest; ACC, accuracy; SEN, sensibility; SPE, speci city; Lasso, Least absolute shrinkage and selection operator; mRMR, Max-Relevance and Min-Redundancy. Figure 1 Work ow diagram. A, three cohorts were enrolled. The cohort 1 (n=183) and 2 (n=51) were both from SILCODE project, and divided qualitatively by Aβ status or future cognitive outcomes, respectively; the cohort 3 (n=37) included 24 "converters" from SILCODE and 13 from ADNI. All participants were cognitively healthy at baseline timepoint and were evaluated in a standardized protocol. B, preprocessing of amyloid-PET and MPMRI was performed for each modality respectively. C, radiomics features were extracted from each modality respectively; a novel method characterized by the combination of function perturbations (t-test, autocorrelation test, and three independent algorithms) and sample perturbations (5-fold cross-validation and 100 repetitions) was performed to select features from the training dataset (80% of cohort 1). D, retained features were incorporated into classi cation models and veri ed in the validation (the rest 20% of cohort 1) and test dataset (cohort 2), respectively. Furthermore, during the process of selection, stable high-frequency features were identi ed, they were undisturbed by perturbations and correlated with the SUVR values, played good roles in predicting prospective cognitive decline (cohort 2). E, levels of stable high-frequency features were tested whether change with cognitive decline or impact the conversion time. Abbreviations: ADNI, Alzheimer's Disease Neuroimaging Initiative; MRI, magnetic resonance imaging; sMRI, structural MRI; fMRI, functional MRI; DTI, diffusion tensor imaging; MPMRI, multi-parameter MRI; PET, positron emission tomography; FA, fractional anisotropy; MD, mean diffusivity; AD, axial diffusivity; RD, radial diffusivity; SUVR, standard uptake value ratio; Aβ, amyloid β; Lasso, Least absolute shrinkage and selection operator; mRMR, Max-Relevance and Min-Redundancy; SVM, support vector machine; RF, random forest; ROC, receiver operating characteristic; AUCs, areas under curve; NC, normal control; CI, cognitive impairment.