Simultaneous Confidence Corridors for neuroimaging data analysis: applications to Alzheimer's Disease diagnosis
Juan A. Arias, Carmen Cadarso-Suárez, Pablo Aguiar-Fernández
SSimultaneous Confidence Corridors forneuroimaging data analysis: applications forAlzheimer’s Disease diagnosis.
Juan A. Arias (cid:66) , Carmen Cadarso-Su´arez , and Pablo Aguiar-Fern´andez Biostatistics and Biomedical Data Science Unit. Department of Statistics,Mathematical Analysis, and Operational Research, Universidade de Santiago deCompostela, Spain Nuclear Medicine Department and Molecular Imaging Group, University ClinicalHospital (CHUS) and Health Research Institute of Santiago de Compostela (IDIS),Santiago de Compostela, Spain Molecular Imaging Group, Department of Psychiatry, Radiology and PublicHealth, Faculty of Medicine, Universidade de Santiago de Compostela, Spain
Abstract.
Alzheimer’s disease (AD) is a chronic neurodegenerative con-dition responsible for most cases of dementia and considered as one ofthe greatest challenges for neuroscience in this century. Early AD signsare usually mistaken for normal age-related cognitive dysfunctions, thuspatients usually start their treatment in advanced AD stages, when itsbenefits are severely limited. AD has no known cure, as such, hope lieson early diagnosis which usually depends on neuroimaging techniquessuch as Positron Emission Tomography (PET). PET data is then ana-lyzed with Statistical Parametric Mapping (SPM) software, which usesmass univariate statistical analysis, inevitably incurring in errors de-rived from this multiple testing approach. Recently, Wang et al. (2019)[1] formulated an alternative: applying functional data analysis (FDA),a relatively new branch of statistics, to calculate mean function and si-multaneous confidence corridors (SCCs) for the difference between twogroups’ PET values. Here we test this approach with a practical appli-cation for AD diagnosis, estimating mean functions and SCCs for thedifference between AD and control group’s PET activity and locatingregions where this difference falls outside estimated SCCs, indicatingdifferences in brain activity attributable to AD-derived neural loss. Ourresults are consistent with previous literature on AD pathology and sug-gest that this FDA approach is more resilient to reductions in sample sizeand less dependent on ad hoc selection of an α level than its counterpart,suggesting that this novel technique is a promising venue for research inthe field of medical imaging. Keywords:
Neurodegenerative disease · Alzheimer’s disease · Medicalimaging · Neuroimage · Positron emission tomography · Functional dataanalysis · Simultaneous confidence corridors · Bivariate splines a r X i v : . [ q - b i o . N C ] J a n Arias et al. (2020) - Working draft
Motivated by Wang et al.’s [1] work, where they propose the use of functionaldata analysis (FDA) to the field of medical imaging, in this article we carry outa practical application by calculating simultaneous confidence corridors (SCCs)of positron emission tomography (PET) data obtained from Alzheimer’s dis-ease (AD) and control (CN) patients. We apply this FDA procedure both forone-sample and two-sample setups, then visually compare results with the onesobtained using Statistical Parametric Mapping (SPM), and assess the potentialutility of this novel technique for AD diagnosis. While a detailed review of ADpathology and the calculation of SCCs is out of scope for the current paper, weaim to give a sufficiently consistent background for all researchers interested inthis subject and to provide computational tools to ensure the replicability of thisstudy.
AD is the most common neurodegenerative disease (NDD) and sole responsiblefor the majority of cases of dementia [2], as well as a major death cause [3]. NDDis an umbrella term which describes a wide number of progressive, incurable, andhighly debilitating medical conditions caused by neural degeneration [4]. NDDare often age-related, as such, vulnerable population to these diseases is expectedto increase substantially in the following decades, with estimates predicting thatpeople over 60 years old will double between the year 2000 and 2050 [5]. By thatyear, patients suffering from dementia will exceed 115 million worldwide [6]. Thisgrowing problematic, known as the burden of neurodegenerative diseases [7], isconsidered one of the greatest challenges for neuroscience in this century, as itthreatens not only patients, but also their families, caregivers, and public healthsystems’ capacities.Risk factors for AD include age, smoking, heart disease, hypercholesterolemia,hypertension, alcohol consumption, low physical activity, and diabetes [8,9,10].On the other hand, factors hypothesized to prevent AD include physical exercise[11], a healthy diet [12], and high cognitive reserve [13], among others [8,10].There are two types of AD depending on onset age: early-onset AD (i.e. familialAD), accounting for 1% of the cases [14]; and late-onset AD (LOAD), account-ing for the remaining 99% of cases. Familial AD causes are genetic and easy todetermine (e.g. the Osaka mutation [15]), but the situation is much less clear forLOAD, with low heritability rates [16] and a variety of 19 loci identified whichare associated with AD but still leaving a large portion of genetic risk withoutexplanation [17].AD main effect is progressive neurological damage resulting from neuronalloss which occurs first in the hippocampus, temporal cortices, and parietal cor-tices [18,19] (see Figure 1). As the disease evolves, neural damage progressesand affects the parahippocampal gyrus and amygdala [20]. In advanced stages,AD affects brain regions involved in primary bodily functions, complicating vi-tal movements and becoming fatal to the patient [2]. Caused by the presence
CCs for Neuroimaging Data Analysis 3 of neuritic plaques [21] and neurofibrillary tangles [22] in the affected neurons,this progressive neurological damage translates into neuropsychological and be-havioural manifestations including worsened capacity to recall information, prob-lems with planning and decision-making, generalized confusion, delusions, hallu-cinations, repetition of conversations, anxiety, among others [23,2]. In addition,AD is typically accompanied by comorbidities such as major depressive disor-ders [24] which complicate the already poor situation of the patient, aggravatingperceived health in a self-sustained downward spiral as proposed by recent life-course models [25] and contributing to severe reductions in life expectancy andother diffuse effects on health and well-being [26]. (a) Coronal view (b) Sagittal left view (c) Axial superior view
Figure 1: Regions of interest for research on early AD diagnosis in the humanbrain: hippocampus (red), temporal cortices (blue), and parietal cortices (green).Despite our knowledge of AD neuropathology, no theoretical framework canexplain the formation of these plaques and tangles, nor the specific regionaldistribution displayed. A number of coexisting and non-exclusive hypothesis havebeen proposed, including the amyloid cascade hypothesis [27], the cholinergichypothesis [28], or the Tau hypothesis [29], among others [30,31].AD diagnosis is truly challenging as the pathological process may start be-tween 20 to 30 years before the first detectable symptoms appear [32,14]. Inaddition, early AD stages are very similar to the common forms of senile demen-tia [33] and mild cognitive impairment (MCI) [34], increasing the complexityof a precise early diagnosis. Several biomarkers have been proposed [23,35], butthey all suffer from limitations and a precise diagnosis usually happens in lateAD stages by means of test batteries such as the Mini Medical State Examina-tion [23] based on standard diagnostic criteria (i.e. DSM-IV-TR) [36]. Anotheroption for AD diagnosis is the use of neuroimaging tools such as magnetic res-onance imaging (MRI) [37] and single-photon emission computed tomography[38]. PET, a neuroimaging technique which measures regional metabolic ratesof radioactively-marked glucose, has been applied to AD diagnosis in early de-velopmental stages [39,40,41,42] and for differential AD diagnosis [43]. A variantof PET, 18F-FDG PET, is the technique used in our study; further details arediscussed in Section 3.1.
Arias et al. (2020) - Working draft
With regards to treatment, current pharmacotherapeutic interventions forAD are carried out with drugs such as cholinesterase inhibitors [44] and me-mantine [45] in order to slow down the progress of the disease. However, resultshave been modest, difficult to translate into generalized medical benefit [23],and limited to only five medical treatments for AD approved by 2016 [46]. Inaddition, these treatments are designed as symptomatic rather than disease-modifying and it is not probable that a single cure for AD will appear [46,47].Non-pharmacoterapeutic treatments are also applied although the evidence ontheir efficiency is incomplete and not lacking contradictions [48].Given the growing problematic of AD in modern societies and the limitedsuccess of therapeutic interventions, there is a growing interest for early AD di-agnosis considering that available treatments are more beneficial in initial stagesof the pathological process [49,50] and that AD-derived neurodegeneration startsbetween 20 to 30 years prior to the appearance of the first clinical symptoms[14,32]. Thus, early diagnosis of AD will be critical to reduce the impact ofthis disease in our societies, especially bearing in mind that - as an age-relatedcondition - small delays on disease onset are expected to markedly reduce ADprevalence [46,51]. Besides, early treatment would slow down or even stabilizethe disease, reducing the need for external care, improving patients’ life quality,and increasing a life expectancy which is currently estimated to range betweenthree to ten years after diagnosis [52]. After this review of the current stateof the art on AD, we consider that early diagnosis is the global imperative forresearchers in order to reduce the impact of this upcoming pandemic.
Having reviewed basic concepts on AD and considering the imperative need forearly diagnosis of this severe condition, the proposed aim of this article is toevaluate potential improvement for AD diagnosis by means of FDA techniques,applying SCCs to the difference between mean functions of PET data from ADand CN groups and examining regions exhibiting differential patterns of activity.In order to achieve our aim we propose the following objectives:(a) To calculate SCCs for one-sample and two-sample cases using bivariatesplines over triangulations of 18FDG-PET data.(b) To visualize areas where the difference between estimated mean func-tions falls outside estimated SCCs.(c) To evaluate whether obtained results are in line with existent scientificliterature on regional distribution of AD-derived neuronal loss.(d) To visually compare these results with the ones obtained using SPMsoftware under the same conditions in order to visually assess perfor-mance of proposed FDA methodology.
CCs for Neuroimaging Data Analysis 5
There are slightly different approaches for PET imaging depending on the ra-dioisotope used; in 18F-FDG PET, Fluorodeoxyglucose (18F-FDG), a radioiso-tope analog of glucose, is used as tracer to monitor brain metabolic rates.Positron emission rates by molecules of 18F-FDG trapped in brain tissues areused as an indirect measure of glucose consumption, which is then reconstructed[53] producing 3D images for the position of this tracer in the brain.18F-FDG PET has been recently included in AD diagnostic routines and isthe methodology chosen for our study. We use Statistical Parametric Mapping(SPM) [54] software to carry out anatomical standardization which allows forinter-subject comparisons after a series of pre-processing stages. This step is keyin any neuroimaging study as the existence of a precise point-to-point correspon-dence is critical when comparing scans from brains which present characteristicshape and size. For this reason, we rely on SPM’s anatomical standardizationroutines as they have been thoroughly developed and tested [55].In order to obtain 18F-FDG PET data we drawn upon the Alzheimer’s Dis-ease Neuroimaging Initiative (ADNI) [56], a platform that collects data fromdifferent research institutions focusing on AD diagnosis. We selected 18F-FDGPET and MRI data, together with demographic information (age and sex) forCN group (75 patients; 44 male; age: 75 . ± .
96 years) and AD group (51 pa-tients; 30 male; age: 74 . ± .
25 years) summing 126 participants. Images werethen realigned, unwrapped, co-registered with MRI data, spatially normalised,mean proportionally scaled, and masked following standard procedures deployedby SPM with the aim of guaranteeing voxel-to-voxel comparability between im-ages (see Figure 2). (a) Raw image (b) Normalized image (c) Masked image
Figure 2: Stages of 18F-FDG PET processing. (a) Raw image obtained fromADNI, (b) image after anatomical standardization, (c) processed image aftermasking.
Arias et al. (2020) - Working draft
The resulting 18F-FDG PET image files were then converted into a regulardata frame for its use in . The result is a data frame where every 18F-FDGPET value is pooled together with its correspondent coordinates in the 3D brainmap and its corresponding demographic information and study group. Detailsfor these bulk transformations using Matlab and can be found and replicatedusing the scripts stored in our open code repository.
In a broad sense, FDA can be defined as the field of statistics which deals withthe theoretical foundations and necessary tools for the analysis of data expressedin the form of functions, which can be extended to data in the form of imagesand shapes. FDA has experimented a great rise in popularity in the last decadesand comprehensive publications on this field - including monographs [57,58] andreview articles [59,60] - are now published explaining FDA’s theoretical basis andapplicability. Recently, Wang and colleagues [1] proposed a novel approach toobtain mean values of imaging data and their corresponding SCCs (also knownas ’simultaneous confidence bands’) using FDA and thus avoiding typical issuesof brain activity estimation such as the leakage in complex data structures andthe multiple comparison problematic. In their article, the authors prove thismethod to be statistically consistent and asymptotically normal under standardregularity conditions.First, images used in biomedical studies are often irregularly shaped, whichimplies a need for reliable smoothing methods. However, such methods usuallysuffer from a problem of leakage with complex data structures such as the oneshere considered, showing difficulties in the estimation for boundary regions. Withour FDA approach, this problem is addressed by applying bivariate splines overDelaunay triangulations [61], thus preserving the features of these complex brainimages [62] (see Figure 3).Following, when analyzing medical imaging data, problems arise not onlyfor the estimation of the mean value for a given point, but also for the estima-tion of associated uncertainty, a problem which gets even more complex whenconsidering the spatial correlation between points. Usually, the predominant an-alytic techniques deployed in software such as SPM follow the mass univariateapproach which considers pixels as independent units and then compares themwith classical methods such as a T -test. The multiple comparisons problem isthen addressed with approaches such as the Bonferroni correction or the randomfield theory [64]. However, these are ad hoc corrections and depend on the choiceof an appropriate threshold, which is based on the designated professional’s ex-pertise. Wang et al.’s proposal considers data as functional, that is, as the resultof a continuously defined function observed on a pre-defined regular grid. Assuch, attention moves from using pixels as basic analytic units to consideringimages as a whole and then calculating corresponding SCCs. The utility of thisapproach for the estimation of confidence bands for functional data has beenpreviously addressed in the scientific literature [65,66]. CCs for Neuroimaging Data Analysis 7(a) N=8 (b) N=15 (c) N=25
Figure 3: Delaunay triangulations for SCCs construction at Z = 30. IncreasingN values represent increases in triangulation’s degree of fineness. Wang et al.[63] consider a fineness degree of N=8 as sufficient.Here we propose the use of this FDA methodology to a practical case usingPET data from CN and AD patients of different sexes and age ranges. Our goalis to show how this method can be applied to neuroimaging data and accuratelypoint out regional differences in neural activity between two groups. We usedlanguage together with packages Triangulation [67], to get custom Delaunaytriangulations for the boundaries of our data, and Wang et al.’s ImageSCC [63]package, to calculate estimated mean functions and corresponding SCCs. Thisanalysis is then replicated filtering by sex and age ranges. Results are shownin the following section, together with a succinct explanation of what can beobserved in the attached figures and a visual comparison using a classic SPMapproach under the same conditions. In this section we describe results obtained using the aforementioned FDA tech-nique to PET data from CN and AD groups. First, in the one-sample case wecalculate estimated mean functions and SCCs for each of the two groups indi-vidually. Then, in the two-sample case, we carry out comparisons between ADand CN entire groups and replicate them filtering by sex and age. These com-parisons determine which brain areas have PET activity levels falling inside ourestimated SCCs for the difference between groups, suggesting normal levels ofbrain activity, and which areas fall outside estimated SCCs for the differencebetween groups, suggesting hypo-activity or hyper-activity patterns. AlthoughPET data is naturally three-dimensional, we use one strategically chosen sliceof our data ( Z = 30) as an exemplar as it cuts across regions of interest for ADdiagnosis such as the hippocampus and temporal cortices [68]. Arias et al. (2020) - Working draft
The first step to interpret our results is to consider the one-sample case. Figure4 shows estimated mean function and SCCs ( α = 0 .
05) for PET activity in CNand AD groups separately, providing a general overview of activity levels in agiven two-dimensional slice. Although informative, these results have a limitedutility as in real practice the aim is to examine differences between treatmentgroups, sexes, age blocks, or a combination of these. An experienced eye couldalready point out differences in activity levels from these images, however, theaim of this article is to provide the necessary tools to accurately detect brainregions which are suffering AD-derived atrophy; to achieve this end we now turnour attention to the two-sample case.
Control Group (a) Lower SCCs (b) Mean Function (c) Upper SCCs
Alzheimer’s Disease Group (d) Lower SCCs (e) Mean Function (f) Upper SCCs
Figure 4: Estimated mean function (center), lower SCCs (left), and upper SCCs(right) for CN group (n=75) and AD group (n=51). SCCs calculated for α = 0 . N = 8). CCs for Neuroimaging Data Analysis 9
For the two-sample case we estimate SCCs for the difference between CN and ADgroups’ mean functions, identify pixels whose PET values fall outside estimatedSCCs (pinpointing abnormal activation patterns), and replicate the analysis forthree different setups: entire group, filtering by sex, and filtering by age. We alsodisplay, as a visual assessment tool, results obtained using SPM under the sameconditions.Bear in mind that this methodology detects both hypo- and hyper-activationpatterns. Finding hyper-activation patterns in AD brain regions compared to CNsounds counter intuitive at first. However, as we further discuss in Section 5, thereare different viable explanations for this finding (e.g., compensatory mechanisms,disturbances in normal brain deactivation processes). Nevertheless, we considerthese findings are more likely attributable to a statistical artifact derived fromthe grand mean scaling and intensity normalization processes. For this reasonwe will only discuss results involving hypo-activity patterns in AD compared toCN groups.
Figure 5 displays obtained results for a comparison between entire CN and ADgroups. Upper row displays results using SPM, while lower row shows resultsobtained using our proposed FDA approach. Results are shown for decreasing α levels in both setups. This figure shows that, for any of the selected α levels, PEThypo-activity is found in temporal regions and hippocampus. This suggests thatthese regions are suffering a process of neural loss attributable to AD pathology,findings which go in line with scientific literature on AD pathology [68]. Bear inmind that other regions which could be relevant for AD diagnosis - such as theparahippocampal gyrus and amygdala - are absent in the observed slice.Results using FDA and SPM show a strong resemblance in terms of regionaldistribution and point towards hypo-activation patterns in the same regions(i.e. temporal and hippocampal areas), although displaying different extensions.Besides, it is remarkable to observe that our FDA methodology appears to bemore resilient to alterations in α levels than its counterpart, displaying resultswhich are similar in distribution and extension across the three selected α levels.In contrast, SPM appears as a methodology whose results heavily depend on theselected α level. Figure 6 shows results for the same comparison between AD and CN groups butnow filtering by sex group. First, we display results for the entire group analysisas a visual reference. Following, we show results for SPM and FDA approachesfiltering by female and male participants, aiming to identify differences in PETactivity between AD and CN groups filtered by sex. In this figure we can see thatobtained results for SPM are much more diffuse than the ones obtained using
Results using SPM (a) α = 0 . α = 0 .
05 (c) α = 0 . Results using FDA (d) α = 0 . α = 0 .
05 (f) α = 0 . Figure 5: Comparison between classical SPM (uncorrected p-value) and FDAapproaches for entire groups: AD (n=51) vs CN (n=75) at different α levels.Blue indicates hypo-activation in AD compared to CN while orange indicateshyper-activation.an FDA approach. SPM results display large regions of hypo-activity both forfemale and male sub-groups which can be identified as parts of the hippocampusand temporal cortices. This noisy response is easily identified, for example, inFigure 6.b where hypo-activity is found in an area that goes from left temporalto left occipital lobes; this result does not go in line with the available literatureon AD pathology and is likely attributable to a poor analysis.On the other hand, results obtained using our FDA approach are more cau-tious and, although the distribution of regions identified as hypo-active remainssimilar to the ones identified with SPM, the extension of these areas is muchmore reduced compared to its counterpart, with hypo-activity nuclei clearly lo-cated in temporal and hippocampal regions both for female and male sub-groupsand identifying a much less numerous amount of sparse nuclei. CCs for Neuroimaging Data Analysis 11
Overall, both methods present results which go in line with the availableliterature (i.e. there are no apparent differences in regional distribution of neuralloss between female and male AD patients) but our FDA approach appears to bemore cautious in the displayed results following reductions in sample size. Thisis easily observed in the current figure where, although regional distributions ofPET hypo-activity remain similar across methodologies, the extension of theseareas is much more constrained in our proposed FDA setup.
Results using SPM (a) Entire group (b) Female sub-group (c) Male sub-group
Results using FDA (d) Entire group (e) Female sub-group (f) Male sub-group
Figure 6: Comparison between SPM ( α = 0 .
05; uncorrected p-value) and FDAapproaches (SCCs for α = 0 .
05) for the analysis of groups filtered by sex. Femalesub-group: CN (n=31) & AD (n=21); male sub-group: CN (n=43) & AD (n=30).Blue indicates hypo-activation in AD compared to CN while orange indicateshyper-activation.
Results using SPM (a) Entire Group (b) ≤
75 y.o. sub-group (c) >
75 y.o. sub-group
Results using FDA (d) Entire group (e) ≤
75 y.o. sub-group (f) >
75 y.o. sub-group
Figure 7: Comparison between SPM ( α = 0 .
05; uncorrected p-value) and FDAapproaches (SCCss for α = 0 .
05) for the analysis of groups filtered by age. ≤
75 y.o. sub-group: CN (n=31) & AD (n=21). >
75 y.o. sub-group: CN (n=43) &AD (n=30). Blue indicates hypo-activation in AD compared to CN while orangeindicates hyper-activation.
Figure 7 shows results for the same comparison between AD and CN groupsfiltering by age block. First, we display results for the entire group analysis as avisual reference. Following, we filter by age, dividing our populations in a ≤ >
75 y.o. sub-group. We perform this analysis in duplicate,both for SPM and for FDA approaches under the same setup. This figure showspatterns of hypo-activation which are similar between SPM and FDA methodolo-gies, pointing towards reductions in PET activity in hippocampal and temporalregions, a pattern which is more easily identified in the ≤
75 y.o. sub-group.SPM results are more diffuse and highlight large regions and numerous nuclei
CCs for Neuroimaging Data Analysis 13 showing hypo-activation. On the other hand, our FDA approach is again morecautious, with reduced extensions and scarcer nuclei identified for the ≤
75 y.o.sub-group. The same happens with the >
75 y.o. sub-group but, in this case, ourFDA approach identifies a reduced number of hypo-activity nuclei where SPMidentifies numerous and scattered nuclei which are difficult to interpret.Further, both methods’ results - although similar, especially for the ≤ ≤
75 y.o. sub-group, but disappear for the >
75 y.o. sub-group. However,it can be argued that differences in brain activity between CN and AD groupsare more intense in early stages of life, as opposed to older populations wherenormal ageing features overlap and create a confusion effect with the properlypathological features of AD. In addition, obtained results might also suffer fromdistortions due to reductions in sample size.
This article focuses on the growing medical problem of AD in our societies, con-sidered as one of the greatest challenges for neuroscience in this century andwith estimates predicting more than a hundred million patients worldwide by2050 [5]. Considering that AD has no known cure and the limited success oftherapeutic interventions, the most promising field of study is early AD diag-nosis. This is supported by research suggesting that AD may begin between 20to 30 years prior to first observable symptoms [14]. Thus, early identificationand treatment appear as a critical step, especially bearing in mind that, as anage-related condition, small delays on disease onset can significantly reduce ADprevalence [46].In this article we intend to improve AD diagnostic tools by comparing classi-cal diagnostic methods with a new proposal whose theoretical basis were recentlysettled by Wang et al. [1]. This new approach avoids the multiple testing cumu-lative errors problem, performs well when dealing with complex data structures,and overall represents a step forward in the way we perform statistical analy-sis of data in the medical imaging field. Our aim for this article is to prove bymeans of visual assessment that results obtained using this novel approach areat least equally valid for AD diagnosis as the ones obtained using traditionalapproaches. Further, we also intend to prove that this methodology might bemore resilient than SPM in analysis that implicate complex structures and bigamounts of data.It is worth noting that our analysis detects both hypo and hyper-activationsin the difference between groups. As a result, although counter intuitive, ourresult figures also display regions with increases in PET activity in AD comparedto CN (e.g., putamen, vermis, cerebellum, some cortical regions). These resultsare not rare in the neuroimaging literature; previous publications have explainedthem suggesting that AD may produce compensatory changes in neural activity[69] or that hyperactivity is actually a sign of disturbance in normal deactivation processes [70]. However, even though there is evidence suggesting the existenceof compensatory mechanisms in AD and disturbances in normal deactivationpatterns, we consider that this effect in our results is more likely attributable toa statistical artifact caused by the process of grand mean scaling and intensitynormalization. For this reason we only evaluate results involving hypo-activitypatterns in AD compared to CN groups.We carried out this process of regional identification for the difference be-tween AD and CN estimated mean functions, first for entire groups and thenfiltering by sex and age blocks. For the entire group analysis, Figure 5 showshypo-activity patterns in temporal and hippocampal regions for any of the se-lected α levels both for traditional SPM and novel FDA approaches. Albeit re-sults were similar, it appears to us that the proposed FDA methodology is morecautious and more resilient to changes in α levels than its SPM counterpart.Besides, Figure 6 and Figure 7 demonstrate that the proposed FDA approachis also more resilient to reductions in sample size than classic SPM approaches,displaying reduced nuclei of hypo-activation both in terms of quantity and size.This is especially easy to visualize in Figure 7.c and Figure 7.f where SPM iden-tifies a big number of sparse and small nuclei showing patterns of hypo-activitywhile, on the other hand, the proposed FDA approach identifies a more reducednumber of brain areas as relevant. Overall, results for both methodologies differin terms of the extension of identified areas, but not in terms of distributionof these areas. This points out towards the validity of our proposed FDA ap-proach, which seems to identify the same areas suffering from AD-derived neuralloss than SPM does.In summary, our goal for this article was to implement Wang et al.’s [1]methodology from a theoretical framework into a practical case in an attempt todemonstrate that this novel methodology is valid for AD diagnosis and poten-tially other medical imaging analysis. After visually assessing results displayedin Figure 5, Figure 6, and Figure 7 we consider our initial goal for this articleas a success and acknowledge this statistical approach’s potential in the fieldof medical imaging. In addition, this FDA approach circumvents the multiplecomparison problem and appears to display a better resistance to changes in α levels and reductions in sample size than classic SPM. However, there is stillmuch work to do in order to refine this methodology and to apply it in routineclinical practice. Although promising, the results obtained using this FDA novel technique sufferfrom intrinsic limitations derived from the proposed setup and available data.We used this methodology with the aim of detecting brain atrophy basing our re-sults on 18FDG-PET data, however, inappropriate use of 18FDG-PET anatomicstandardization processes can cause a series of statistical artifacts and system-atic biases which have to be taken in account to avoid a serious problem of falsepositives [71,72].
CCs for Neuroimaging Data Analysis 15
Besides, this technique has been applied to a limited number of participants(126 participants: 51 AD & 75 CN). Future research should increase the num-ber of participants and also increase the number of conditions considered byincluding transitional stages between healthy and AD populations such as pa-tients suffering from MCI. Additionally, as we stated previously in this article,neuroimaging data is naturally three-dimensional. In this article we work witha selected slice of two-dimensional data which is especially relevant as severalbrain regions involved in AD diagnosis cut through it. However, future researchshould search for a way of extending this methodology to the three-dimensionalcase without losing estimator’s consistency. This would be a great step forwardfor neuroscience and medical imaging in general, albeit not without lots of effortboth in theoretical and computational terms.It would also be of great interest to implement this technique into a single-participant setup, as in clinical practice the most useful approach for diagnosisconsists on comparing one single patient with data from a large group of CN pa-tients in order to visualize which brain areas show different patterns of activationand thus decide whether the patient suffers from AD or not. However, the appli-cation of this methodology to a single-case setup has been so far unsuccessful.In addition, the herein proposed methodology would provide a new tool for clin-icians to make grounded diagnosis. However, future research and computationaleffort should strive for the implementation of an automatic classification systembased on SCCs for neuroimaging data. Such a system would not only providethe clinical specialist with better images to make a decision but also with anautomatic proposed classification (either as healthy or pathological) based onlarge datasets of previous patients.Furthermore, although the utility of this technique is confirmed visually bythe great resemblance between considered approaches, future research shouldstrive for fully establishing SCCs as an alternative to traditional methodologiesby mathematically demonstrating its performance compared to SPM in differentsetups. Finally, as we mentioned before in this article, the proposed methodol-ogy is not limited to AD research but rather the opposite, as it is potentiallyextensible to any other neurological diseases and, overall, to any other projectinvolving medical imaging in which neural activity and/or volume calculationsare of clinical interest. Future research should aim for the implementation of thistechnique to other research projects in order to fully elucidate the potential ofthis novel approach.
Acknowledgments
The authors are grateful for the collaboration and input provided by severalmembers of the GRID-BDS team. Neuroimaging data was extracted from theADNI database which is curated by a series of researchers to whom, even thoughthey did not participate in the elaboration of this research paper, we are verygrateful. Juan A. Arias also wants to acknowledge the passive support of hisradio partner @ uberflut . May the red light bulb shine forever in their studio.
References
1. Wang, Y., Wang, G., Wang, L., Ogden, R.T.: Simultaneous confidence corridors formean functions in functional data analysis of imaging data. Biometrics (2), 427–437 (jun 2020). https://doi.org/10.1111/biom.13156, https://onlinelibrary.wiley.com/doi/abs/10.1111/biom.131562. Association, A., et al.: 2018 alzheimer’s disease facts and figures. Alzheimer’s &Dementia (3), 367–429 (2018)3. Helmer, C., Joly, P., Letenneur, L., Commenges, D., Dartigues, J.F.: Mor-tality with Dementia: Results from a French Prospective Community-basedCohort. American Journal of Epidemiology (7), 642–648 (oct 2001).https://doi.org/10.1093/aje/154.7.642, https://doi.org/10.1093/aje/154.7.6424. Przedborski, S., Vila, M., Jackson-Lewis, V.: Neurodegeneration: What is itand where are we? Journal of Clinical Investigation (10), 1252–1265 (oct 2019). https://doi.org/10.1111/ene.13972, https://onlinelibrary.wiley.com/doi/abs/10.1111/ene.139728. Patterson, C., Feightner, J.W., Garcia, A., Hsiung, G.Y.R., MacKnight, C.,Sadovnick, A.D.: Diagnosis and treatment of dementia: 1. risk assessment andprimary prevention of alzheimer disease. Cmaj (5), 548–556 (2008)9. Rosendorff, C., Beeri, M.S., Silverman, J.M.: Cardiovascular Risk Factors forAlzheimer’s Disease. The American Journal of Geriatric Cardiology (3), 143–149(mar 2007). https://doi.org/10.1111/j.1076-7460.2007.06696.x, http://doi.wiley.com/10.1111/j.1076-7460.2007.06696.x10. Mayeux, R., Stern, Y.: Epidemiology of Alzheimer Disease. Cold SpringHarbor Perspectives in Medicine (8), a006239–a006239 (aug 2012).https://doi.org/10.1101/cshperspect.a006239, http://perspectivesinmedicine.cshlp.org/lookup/doi/10.1101/cshperspect.a006239CCs for Neuroimaging Data Analysis 1711. Podewils, L.J.: Physical Activity, APOE Genotype, and Dementia Risk: Find-ings from the Cardiovascular Health Cognition Study. American Journal of Epi-demiology (7), 639–651 (apr 2005). https://doi.org/10.1093/aje/kwi092, https://academic.oup.com/aje/article-lookup/doi/10.1093/aje/kwi09212. Scarmeas, N., Stern, Y., Mayeux, R., Luchsinger, J.A.: Mediterranean Diet,Alzheimer Disease, and Vascular Mediation. Archives of Neurology (12), 1709(dec 2006). https://doi.org/10.1001/archneur.63.12.noc60109, http://archneur.jamanetwork.com/article.aspx?doi=10.1001/archneur.63.12.noc6010913. Xu, H., Yang, R., Qi, X., Dintica, C., Song, R., Bennett, D.A., Xu, W.: Associationof lifespan cognitive reserve indicator with dementia risk in the presence of brainpathologies. Jama Neurology (10), 1184–1191 (2019)14. Goedert, M., Spillantini, M.G.: A Century of Alzheimer’s Disease. Science { & } doi=10.3233/JAD-2010-10151517. Lambert, J.C., Ibrahim-Verbaas, C.A., Harold, D., Naj, A.C., Sims, R., Bellenguez,C., Jun, G., DeStefano, A.L., Bis, J.C., Beecham, G.W., Grenier-Boley, B., Russo,G., Thornton-Wells, T.A., Jones, N., Smith, A.V., Chouraki, V., Thomas, C.,Ikram, M.A., Zelenika, D., Vardarajan, B.N., Kamatani, Y., Lin, C.F., Gerrish,A., Schmidt, H., Kunkle, B., Dunstan, M.L., Ruiz, A., Bihoreau, M.T., Choi, S.H.,Reitz, C., Pasquier, F., Hollingworth, P., Ramirez, A., Hanon, O., Fitzpatrick, A.L.,Buxbaum, J.D., Campion, D., Crane, P.K., Baldwin, C., Becker, T., Gudnason, V.,Cruchaga, C., Craig, D., Amin, N., Berr, C., Lopez, O.L., De Jager, P.L., Derame-court, V., Johnston, J.A., Evans, D., Lovestone, S., Letenneur, L., Mor´on, F.J.,Rubinsztein, D.C., Eiriksdottir, G., Sleegers, K., Goate, A.M., Fi´evet, N., Huentel-man, M.J., Gill, M., Brown, K., Kamboh, M.I., Keller, L., Barberger-Gateau, P.,McGuinness, B., Larson, E.B., Green, R., Myers, A.J., Dufouil, C., Todd, S., Wal-lon, D., Love, S., Rogaeva, E., Gallacher, J., St George-Hyslop, P., Clarimon, J.,Lleo, A., Bayer, A., Tsuang, D.W., Yu, L., Tsolaki, M., Boss`u, P., Spalletta, G.,Proitsi, P., Collinge, J., Sorbi, S., Sanchez-Garcia, F., Fox, N.C., Hardy, J., Naranjo,M.C.D., Bosco, P., Clarke, R., Brayne, C., Galimberti, D., Mancuso, M., Matthews,F., Moebus, S., Mecocci, P., Del Zompo, M., Maier, W., Hampel, H., Pilotto, A.,Bullido, M., Panza, F., Caffarra, P., Nacmias, B., Gilbert, J.R., Mayhaus, M., Lan-nfelt, L., Hakonarson, H., Pichler, S., Carrasquillo, M.M., Ingelsson, M., Beekly,D., Alvarez, V., Zou, F., Valladares, O., Younkin, S.G., Coto, E., Hamilton-Nelson,K.L., Gu, W., Razquin, C., Pastor, P., Mateo, I., Owen, M.J., Faber, K.M., Jons-son, P.V., Combarros, O., O’Donovan, M.C., Cantwell, L.B., Soininen, H., Blacker,D., Mead, S., Mosley, T.H., Bennett, D.A., Harris, T.B., Fratiglioni, L., Holmes, C.,de Bruijn, R.F.A.G., Passmore, P., Montine, T.J., Bettens, K., Rotter, J.I., Brice,A., Morgan, K., Foroud, T.M., Kukull, W.A., Hannequin, D., Powell, J.F., Nalls,8 Arias et al. (2020) - Working draftM.A., Ritchie, K., Lunetta, K.L., Kauwe, J.S.K., Boerwinkle, E., Riemenschneider,M., Boada, M., Hiltunen, M., Martin, E.R., Schmidt, R., Rujescu, D., Wang, L.S.,Dartigues, J.F., Mayeux, R., Tzourio, C., Hofman, A., N¨othen, M.M., Graff, C.,Psaty, B.M., Jones, L., Haines, J.L., Holmans, P.A., Lathrop, M., Pericak-Vance,M.A., Launer, L.J., Farrer, L.A., van Duijn, C.M., Van Broeckhoven, C., Moskv-ina, V., Seshadri, S., Williams, J., Schellenberg, G.D., Amouyel, P.: Meta-analysisof 74,046 individuals identifies 11 new susceptibility loci for Alzheimer’s disease.Nature Genetics (7), 939 (1984)19. Wenk, G.L.: Neuropathologic changes in Alzheimer’s disease. Journal of ClinicalPsychiatry , 7–10 (2003)20. Mahanand, B., Suresh, S., Sundararajan, N., Aswatha Kumar, M.: Identi-fication of brain regions responsible for Alzheimer’s disease using a Self-adaptive Resource Allocation Network. Neural Networks , 313–322 (aug2012). https://doi.org/10.1016/j.neunet.2012.02.035, https://linkinghub.elsevier.com/retrieve/pii/S089360801200064021. Iwatsubo, T., Odaka, A., Suzuki, N., Mizusawa, H., Nukina, N., Ihara, Y.: Visual-ization of A β
42 (43) and A β
40 in senile plaques with end-specific A β monoclonals:evidence that an initially deposited species is A β
42 (43). Neuron (1), 45–53(1994)22. Clark, C.M., Ewbank, D., Lee, V.M.Y., Trojanowski, J.Q.: Molecular pathology ofAlzheimer’s disease: neuronal cytoskeletal abnormalities. Blue books of practicalneurology , 285–304 (1998)23. Jalbert, J.J., Daiello, L.A., Lapane, K.L.: Dementia of theAlzheimer Type. Epidemiologic Reviews (1), 15–34 (may 2008).https://doi.org/10.1093/epirev/mxn008, https://academic.oup.com/epirev/article-lookup/doi/10.1093/epirev/mxn00824. Lyketsos, C.G., Olin, J.: Depression in Alzheimer’s disease: overviewand treatment. Biological Psychiatry (3), 243–252 (aug 2002).https://doi.org/10.1016/S0006-3223(02)01348-3, http://linkinghub.elsevier.com/retrieve/pii/S000632230201348325. Kemp, A.H., Arias, J.A., Fisher, Z.: Social ties, health and wellbeing: a litera-ture review and model. In: Neuroscience and Social Science, pp. 397–427. Springer(2017)26. Arias, J.A., Williams, C., Raghvani, R., Aghajani, M., Baez, S., Belzung, C.,Booij, L., Busatto, G., Chiarella, J., Fu, C.H., Ibanez, A., Liddell, B.J., Lowe,L., Penninx, B.W., Rosa, P., Kemp, A.H.: The neuroscience of sadness: A multi-disciplinary synthesis and collaborative review. Neuroscience & Biobehavioral Re-views , 199–228 (apr 2020). https://doi.org/10.1016/j.neubiorev.2020.01.006,https://linkinghub.elsevier.com/retrieve/pii/S014976341830614627. Hardy, J.: Alzheimer’s disease: the amyloid cascade hypothesis: an update andreappraisal. Journal of Alzheimer’s disease (s3), 151–153 (2006)28. Francis, P.T., Palmer, A.M., Snape, M., Wilcock, G.K.: The cholinergic hypothesisof Alzheimer’s disease: a review of progress. Journal of Neurology, Neurosurgery& Psychiatry (2), 137–147 (feb 1999). https://doi.org/10.1136/jnnp.66.2.137,http://jnnp.bmj.com/cgi/doi/10.1136/jnnp.66.2.137CCs for Neuroimaging Data Analysis 1929. Maccioni, R.B., Far´ıas, G., Morales, I., Navarrete, L.: The Revitalized Tau Hypoth-esis on Alzheimer’s Disease. Archives of Medical Research (3), 226–231 (apr2010). https://doi.org/10.1016/j.arcmed.2010.03.007, https://linkinghub.elsevier.com/retrieve/pii/S018844091000050030. Moulton, P.V., Yang, W.: Air Pollution, Oxidative Stress, and Alzheimer’sDisease. Journal of Environmental and Public Health (3), 608–622 (nov2014). https://doi.org/10.1016/j.neuron.2014.10.038, https://linkinghub.elsevier.com/retrieve/pii/S089662731400958133. Terry, R.D., Davies, P.: Dementia of the Alzheimer Type.Annual Review of Neuroscience { & } doi=10.3233/JAD-14084336. American Psychiatric Association: Diagnostic and statistical manual of mentaldisorders (DSM-5). American Psychiatric Pub (2013)37. Jack, C.R., Dickson, D.W., Parisi, J.E., Xu, Y.C., Cha, R.H., O’brien, P.C., Ed-land, S.D., Smith, G.E., Boeve, B.F., Tangalos, E.G.: Antemortem MRI findingscorrelate with hippocampal neuropathology in typical aging and dementia. Neu-rology (5), 750–757 (2002)38. McNeill, R., Sare, G.M., Manoharan, M., Testa, H.J., Mann, D.M.A., Neary, D.,Snowden, J.S., Varma, A.R.: Accuracy of single-photon emission computed tomog-raphy in differentiating frontotemporal dementia from Alzheimer’s disease. Journalof Neurology, Neurosurgery & Psychiatry (4), 350–355 (2007)39. Silverman, D.H.S., Gambhir, S.S., Huang, H.W.C., Schwimmer, J., Kim, S., Small,G.W., Chodosh, J., Czernin, J., Phelps, M.E.: Evaluating early dementia withand without assessment of regional cerebral metabolism by PET: a comparison ofpredicted costs and benefits. Journal of Nuclear Medicine (2), 253–266 (2002)40. Mosconi, L.: Glucose metabolism in normal aging and Alzheimer’s disease: method-ological and physiological considerations for PET studies. Clinical and Trans-lational Imaging (4), 217–233 (aug 2013). https://doi.org/10.1007/s40336-013-0026-y, http://link.springer.com/10.1007/s40336-013-0026-y41. Marcus, C., Mena, E., Subramaniam, R.M.: Brain PET in the Diagnosis ofAlzheimer’s Disease. Clinical Nuclear Medicine (10), e413–e426 (oct 2014).https://doi.org/10.1097/RLU.0000000000000547, http://content.wkhealth.com/linkback/openurl?sid=WKPTLP:landingpage { & } an=00003072-201410000-000260 Arias et al. (2020) - Working draft42. Nordberg, A., Rinne, J.O., Kadir, A., L˚angstr¨om, B.: The use of PETin Alzheimer disease. Nature Reviews Neurology (5), 685–691 (2004)44. Birks, J.S.: Cholinesterase inhibitors for Alzheimer’s disease. In: Birks, J.S. (ed.)Cochrane Database of Systematic Reviews. John Wiley & Sons, Ltd, Chich-ester, UK (jan 2006). https://doi.org/10.1002/14651858.CD005593, http://doi.wiley.com/10.1002/14651858.CD00559345. Robinson, D.M., Keating, G.M.: Memantine. Drugs (11), 1515–1534 (2006)46. Briggs, R., Kennelly, S.P., O’Neill, D.: Drug treatments inAlzheimer’s disease. Clinical Medicine (7), 702–716 (jul 2010). https://doi.org/10.1016/S1474-4422(10)70119-8, https://linkinghub.elsevier.com/retrieve/pii/S147444221070119848. Dietch, J.T., Hewett, L.J., Jones, S.: Adverse effects of reality orientation. Journalof the American Geriatrics Society (10), 974–976 (1989)49. Cummings, J.L., Doody, R., Clark, C.: Disease-modifying therapies for Alzheimerdisease: Challenges to early intervention. Neurology (3), 186–191 (jul2007). https://doi.org/10.1016/j.jalz.2007.04.381, http://doi.wiley.com/10.1016/j.jalz.2007.04.38152. Zanetti, O., Solerte, S., Cantoni, F.: Life expectancy in Alzheimer’s Dis-ease (AD). Archives of Gerontology and Geriatrics , 237–243 (jan2009). https://doi.org/10.1016/j.archger.2009.09.035, https://linkinghub.elsevier.com/retrieve/pii/S016749430900235053. Leahy, R.M., Qi, J.: Statistical approaches in quantitative positron emission to-mography. Statistics and Computing (2), 147–165 (2000)54. Penny, W.D., Friston, K.J., Ashburner, J.T., Kiebel, S.J., Nichols, T.E.: Statisticalparametric mapping: the analysis of functional brain images. Elsevier (2011)55. Ishii, K., Willoch, F., Minoshima, S., Drzezga, A., Ficaro, E.P., Cross, D.J., Kuhl,D.E., Schwaiger, M.: Statistical brain mapping of 18F-FDG PET in Alzheimer’sdisease: validation of anatomic standardization for atrophied brains. Journal ofNuclear Medicine (4), 548–557 (2001)56. Mueller, S.G., Weiner, M.W., Thal, L.J., Petersen, R.C., Jack, C.R.,Jagust, W., Trojanowski, J.Q., Toga, A.W., Beckett, L.: Ways towardan early diagnosis in Alzheimer’s disease: The Alzheimer’s Disease Neu-roimaging Initiative (ADNI). Alzheimer’s & Dementia (1), 55–66 (julCCs for Neuroimaging Data Analysis 212005). https://doi.org/10.1016/j.jalz.2005.06.003, http://doi.wiley.com/10.1016/j.jalz.2005.06.00357. Ramsay, J., Silverman, B.: Principal components analysis for functional data. Func-tional data analysis pp. 147–172 (2005)58. Ferraty, F., Vieu, P.: Nonparametric functional data analysis: theory and practice.Springer Science & Business Media (2006)59. Wang, J.L., Chiou, J.M., M¨uller, H.G.: Functional data analysis. Annual Reviewof Statistics and Its Application , 257–295 (2016)60. Ullah, S., Finch, C.F.: Applications of functional data analysis: A systematic re-view. BMC medical research methodology (1), 43 (2013)61. Chew, L.P.: Constrained delaunay triangulations. Algorithmica (1-4), 97–108(1989)62. Lai, M.J., Wang, L.: Bivariate penalized splines for regression. Statistica Sinica pp.1399–1417 (2013)63. Wang, Y., Wang, G., Wang, L.: ImageSCC: SCC for Mean Function of ImagingData (2020), r package version 0.1.064. Worsley, K.J., Taylor, J.E., Tomaiuolo, F., Lerch, J.: Unified univari-ate and multivariate random field theory. NeuroImage , S189–S195(jan 2004). https://doi.org/10.1016/j.neuroimage.2004.07.026, https://linkinghub.elsevier.com/retrieve/pii/S105381190400398265. Degras, D.A.: Simultaneous confidence bands for nonparametric regression withfunctional data. Statistica Sinica pp. 1735–1765 (2011)66. Choi, H., Reimherr, M.: A geometric approach to confidence regions and bands forfunctional parameters. Journal of the Royal Statistical Society: Series B (StatisticalMethodology) (1), 239–260 (2018)67. Lai, M.J., Wang, L.: Triangulation: Triangulation in 2D domain (2020), r packageversion 0.1.068. da Silva Filho, S.R.B., Barbosa, J.H.O., Rondinoni, C., dos Santos, A.C., Salmon,C.E.G., da Costa Lima, N.K., Ferriolli, E., Moriguti, J.C.: Neuro-degenerationprofile of alzheimer’s patients: A brain morphometry study. NeuroImage: Clinical , 15–24 (2017)69. Schwindt, G.C., Black, S.E.: Functional imaging studies of episodic memory inalzheimer’s disease: a quantitative meta-analysis. Neuroimage (1), 181–190(2009)70. Bejanin, A., Viard, A., Ch´etelat, G., Clarys, D., Bernard, F., P´elerin, A.,de La Sayette, V., Eustache, F., Desgranges, B.: When higher activations reflectlower deactivations: a pet study in alzheimer’s disease during encoding and retrievalin episodic memory. Frontiers in human neuroscience , 107 (2012)71. Ishii, K., Willoch, F., Minoshima, S., Drzezga, A., Ficaro, E.P., Cross, D.J., Kuhl,D.E., Schwaiger, M.: Statistical brain mapping of 18f-fdg pet in alzheimer’s disease:validation of anatomic standardization for atrophied brains. Journal of NuclearMedicine (4), 548–557 (2001)72. L´opez-Gonz´alez, F.J., Silva-Rodr´ıguez, J., Paredes-Pacheco, J., Ni˜nerola-Baiz´an, A., Efthimiou, N., Mart´ın-Mart´ın, C., Moscoso, A., Ruibal,´A., Ro´e-Vellv´e, N., Aguiar, P.: Intensity normalization methods inbrain FDG-PET quantification. NeuroImage222