Classifying Alzheimer’s Disease using MRIs and Transcriptomic Data

∗

Lucia Maddalena

1 a

, Ilaria Granata

1 b

, Maurizio Giordano, Mario Manzo

2 c

Mario Rosario Guarracino

3 d

and Alzheimer’s Disease Neuroimaging Initiative (ADNI)

∗

Inst. for High-Performance Computing and Networking, National Research Council,Via P. Castellino, 111, Naples, Italy

Information Technology Services, University of Naples “L’Orientale”,Via Nuova Marina, 59, Naples, Italy

University of Cassino and Southern Lazio, Cassino, Italy

Keywords:

Data Integration, Alzheimers’ Disease, Omics Imaging, Transcriptomics, Magnetic Resonance Imaging.

Abstract:

Early diagnosis of neurodegenerative diseases is essential for the effectiveness of treatments to delay the onset

of related symptoms. Our focus is on methods to aid in diagnosing Alzheimer’s disease, the most widespread

neurocognitive disorder, that rely on data acquired by non-invasive techniques and that are compatible with

the limitations imposed by pandemic situations. Here, we propose integrating multi-modal data consisting of

omics (gene expression values extracted by blood samples) and imaging (magnetic resonance images) data,

both available for some patients in the Alzheimer’s Disease Neuroimaging Initiative dataset. We show how

a suitable integration of omics and imaging data, using well-known machine learning techniques, can lead to

better classiﬁcation results than any of them taken separately, also achieving performance competitive with the

state-of-the-art.

1 INTRODUCTION

Dementia is a public health problem that affects about

50 million people in the world (WHO, 2019). It

is growing rapidly, counting around 10 million new

cases worldwide each year, with an estimate that this

number will triple by 2050. Dementia manifests itself

with a cognitive decline of the patient leading to the

inability to carry out daily life activities (Birkenbihl

et al., 2020). In addition to devastating the lives of pa-

tients and their families, this disease has a signiﬁcant

economic burden on society, estimated at around 600

billion $ per year in 2013 (Birkenbihl et al., 2020) and

https://orcid.org/0000-0002-0567-4624

https://orcid.org/0000-0002-3450-4667

https://orcid.org/0000-0001-8727-9865

https://orcid.org/0000-0003-2870-8134

∗

The Alzheimer’s Disease Neuroimaging Initia-

tive: Data used in preparation of this article were

obtained from the Alzheimer’s Disease Neuroimag-

ing Initiative (ADNI) database (adni.loni.usc.edu). As

such, the investigators within the ADNI contributed

to the design and implementation of ADNI and/or

provided data but did not participate in analysis or

writing of this report. A complete listing of ADNI

investigators can be found at: http://adni.loni.usc.edu/wp-

content/uploads/how to apply/ADNI Acknowledgement

List.pdf

expected to reach around 2 trillion $ per year in 2030

(WHO, 2019). The most common of the forms of

dementia, Alzheimer’s disease (AD), is a progressive

disease whose pathology begins years before the cog-

nitive symptoms appear and are diagnosed by the clin-

ician. Early intervention, in the pre-symptomatic and

not cognitively disabling stages of the disease, is in-

strumental in any future therapy aimed at treating the

disease (Birkenbihl et al., 2021). Indeed, the effec-

tiveness of the treatment often depends on the stage

of the disease. For example, dietary supplements of

folic acid and vitamin B have been shown to improve

cognitive deﬁcits in patients with mild AD, while they

are of little beneﬁt to patients with severe AD (Lee

and Lee, 2020). However, early intervention poses

the problem of diagnosing a patient with AD before

the cognitive symptoms indicate the presence of the

disease itself. An approach to this problem is based

on the analysis of informative biomarkers of the dis-

ease, whose discovery and validation are possible by

having large sets of data available (Birkenbihl et al.,

2021).

In recent years, several longitudinal studies

(groups of patients followed over time in a set of

consecutive specialist investigations) have been con-

ducted to identify biomarkers for the early diagno-

sis of AD and to estimate progression from the in-

Maddalena, L., Granata, I., Giordano, M., Manzo, M., Guarracino, M. and Alzheimer’s Disease Neuroimaging Initiative (ADNI), .

Classifying Alzheimer’s Disease using MRIs and Transcriptomic Data.

DOI: 10.5220/0010902900003123

In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - Volume 2: BIOIMAGING, pages 70-79

ISBN: 978-989-758-552-4; ISSN: 2184-4305

termediate state of the disease (Mild Cognitive Im-

pairment, MCI) to the ﬁnal AD stage, or possibly

to its regression (Cognitively Normal, CN) (Birken-

bihl et al., 2021; Lovestone et al., 2009; Mueller

et al., 2005). The related datasets collect differ-

ent data modalities, which generally include demo-

graphic variables (age, sex, education, etc.), clinical

evaluation (results of cognitive tests, such as MMSE

- Mini Mental State Examination; CDR-SB - Clinical

Dementia Rating Scale Sum of Boxes; AD Assess-

ment Scale ADAS-Cog11 and ADAS-Cog13), geno-

type (status of APOE4 - the apolipoprotein-e4 gene

which represents the major risk factor for AD; sin-

gle nucleotide polymorphisms - SNP - associated

with AD) and magnetic resonance imaging (MRI),

to quantify the atrophy of areas of the brain from

volumes, cortical thickness and surface areas. More

recently, further imaging modalities are being made

available, including Positron Emission Tomography

(PET) with FDG-ﬂuorodeoxyglucose, which mea-

sures cell metabolism, or Diffusion Tensor Imag-

ing (DTI), for estimation of microstructural param-

eters related to cells and axons. Other modalities

sometimes considered include measurements of cere-

brospinal ﬂuid (CSF), to estimate the levels of the pro-

tein markers beta-amyloid, tau and phosphorylated

tau, or transcriptomics data, such as gene expression

(GE) values extracted from biopsy/autopsy or blood.

Much of the studies based on existing longitudi-

nal datasets are devoted to predicting the progression

of the disease over time, as illustrated in recent sur-

veys (Lawrence et al., 2017; Mart

ı-Juan et al., 2020).

Other research is aimed at the diagnosis of each pa-

tient, to classify the degree of disease (CN, MCI or

AD) based on the results of a predetermined visit. Ex-

amples include methods based on various omics data,

such as GE data (Lee and Lee, 2020; Li et al., 2018;

Voyle et al., 2016), or imaging data, such as MRIs and

PETs (Aderghal et al., 2017; Aderghal et al., 2018;

ackstr

om et al., 2018; Li and Liu, 2018; Shi et al.,

2018; Bae et al., 2020). Some research started fo-

cusing on the integration of omics data with infor-

mation from bio-medical images (Nho et al., 2016;

Peng et al., 2016; Maddalena et al., 2020; Maddalena

et al., 2021). Bringing together information coming

from different sources, these omics imaging meth-

ods (Antonelli et al., 2019) can lead to revealing hid-

den genotype-phenotype relationships, with the aim

of better understanding the onset and progression of

many diseases and identifying new diagnostic and

prognostic biomarkers.

Our research aims to develop methods for the clas-

siﬁcation of patients potentially affected by AD that

are helpful for the clinical diagnosis of the disease and

that exploit multi-modal information on the patient’s

status that is readily available. In view of the current

pandemic, which limits the possibility of patient ac-

cess to dedicated and highly specialized medical in-

frastructures, here we are interested in the early diag-

nosis of AD based on the results of individual exam

sessions rather than on longitudinal studies. Further-

more, we focus on those multi-modal data that can be

collected through easily accessible and not extremely

invasive procedures (such as blood tests and MRIs,

see Fig. 1), thus excluding, for example, those com-

ing from brain tissue or CSF.

Figure 1: Omics imaging data adopted for the experiments.

We show how a suitable integration of imaging

and omics features can lead to better results than

any of them taken separately. The proposed ap-

proach, based on existing machine learning tech-

niques, achieves accuracy performance competitive

with state-of-the-art methods, often based on deep

learning.

The paper is organized as follows. Section 2 ex-

plains the proposed method, describing the extrac-

tion procedure for both types of features, imaging

and omics, adopted. Section 3 discusses the results

achieved with the proposed framework and compares

them with those obtained with state-of-the-art ap-

proaches. Finally, Section 4 concludes our paper and

gives some future research directions.

2 MATERIAL AND METHODS

Data used in the preparation of this article were ob-

tained from the Alzheimer’s Disease Neuroimaging

Initiative (ADNI) database (adni.loni.usc.edu). The

ADNI was launched in 2003 as a public-private part-

nership, led by Principal Investigator Michael W.

Weiner, MD. The primary goal of ADNI has been to

test whether serial MRI, PET, other biological mark-

ers, and clinical and neuropsychological assessment

can be combined to measure the progression of MCI

Classifying Alzheimer’s Disease using MRIs and Transcriptomic Data

and early AD.

As our intent was to consider multi-modal infor-

mation integrating omics and imaging data acquired

by non-invasive techniques, we considered gene ex-

pression values extracted by blood samples and MRIs.

GE data from blood samples, collected between 2010

and 2012, is available for 744 ADNI patients. Only

720 of them also have T1-weighted MRIs in the same

time period. According to the baseline visit, 42 pa-

tients have been classiﬁed as AD, 428 as MCI, and

250 as CN, as summarized in Table 1.

Table 1: Classiﬁcation of the selected subset of 720 ADNI

patients for which both GE and MRI data are available.

AD MCI CN

42 428 250

2.1 Extraction of Imaging Features

Imaging features from MRIs have been extracted us-

ing an open-source framework for reproducible evalu-

ation of AD classiﬁcation using conventional machine

learning methods (Samper-Gonz

alez et al., 2018), re-

cently extended to include deep learning CNN-based

methods (Wen et al., 2020), named ClinicaDL. The

framework comprises i) tools to automatically convert

three publicly available datasets, including ADNI,

into the Brain Imaging Data Structure (BIDS) format

(Gorgolewski et al., 2016) and ii) a modular set of pre-

processing pipelines, feature extraction and classiﬁca-

tion methods, together with an evaluation framework,

that provide a baseline for benchmarking the different

components. Its extension includes a modular set of

image preprocessing procedures, CNN classiﬁcation

architectures, and evaluation procedures dedicated to

deep learning. The benchmarking presented in (Wen

et al., 2020) shows that various 3D CNN approaches

achieved similar performances, higher than those of

the 2D slice approach, but still comparable to those

achieved via Support Vector Machine (SVM) (Vap-

nik, 1995) using voxel-based features. Therefore, we

adopted the framework to generate the voxel-based

features from MRIs.

The ADNI MRI data have been curated and con-

verted to the BIDS format using Clinica (Routier

et al., 2021; Samper-Gonz

alez et al., 2018). Then the

T1-volume pipeline of Clinica was adopted, which

is a wrapper of the Segmentation, Run Dartel, and

Normalise to MNI (Montreal Neurological Institute)

Space routines implemented in the Statistical Para-

metric Mapping (SPM, https://www.ﬁl.ion.ucl.ac.uk/

spm/) package. First, the Uniﬁed Segmentation pro-

cedure (Ashburner and Friston, 2005) is adopted to

simultaneously perform tissue segmentation, bias cor-

rection, and spatial normalization of each input im-

age. Next, a group template is created using DAR-

TEL, an algorithm for diffeomorphic image regis-

tration (Ashburner, 2007), from the subjects’ tissue

probability maps on the native space, obtained at the

previous step. The DARTEL to MNI method (Ash-

burner, 2007) is then applied, providing the registra-

tion of the native space images into the MNI space.

As a result, all the images are in a common space, pro-

viding a voxel-wise correspondence across subjects.

A set of imaging features is extracted based on re-

gional measurements, where the anatomical regions

are obtained by an atlas in MNI space, and the av-

erage gray matter density is computed in each of the

regions. In the experiments, the AICHA (Joliot et al.,

2015) atlas (providing 385 regional features) has been

chosen as reference atlas for the AD vs. CN and AD

vs. MCI tasks, while the AAL2 (Tzourio-Mazoyer

et al., 2002) atlas (providing 121 regional features)

for the MCI vs. CN task, as they lead to highest clas-

siﬁcation performance. A simpliﬁed scheme of the

image feature extraction process is reported in Fig.

2. For each of the 720 patients, these imaging fea-

tures have been extracted by the MRIs coming from

the visit closest in time to that of the corresponding

GE data sample.

Figure 2: Extraction of imaging features.

2.2 Extraction of Omics Features

Normalized gene expression proﬁling data from

blood samples of ADNI participants, produced by

Affymetrix Human Genome U219 Array (Affymetrix

(www.affymetrix.com), Santa Clara, CA), were

downloaded from the ADNI website. The dataset

contained 49386 probes. Multiple probes correspond-

ing to the same gene identiﬁer were aggregated by

median value. Signiﬁcance Analysis of Microarrays

(SAM) (Tusher et al., 2001), in the form of R pack-

age, was used for ﬁnding signiﬁcant differentially ex-

pressed genes (DEGs) from the three different un-

BIOIMAGING 2022 - 9th International Conference on Bioimaging

paired two-class comparisons (AD vs. CN, AD vs.

MCI, MCI vs. CN). Both standard (t-statistic) and

Wilcoxon tests were used, random seed generated,

100 Permutations and Delta slider set. Genes were

considered differentially expressed if the q-value was

less than 5%. The best performance results have

been obtained with the features extracted using the

Wilcoxon test for the AD vs. CN (181 features) and

AD vs. MCI (211 features) classiﬁcation tasks. Re-

garding the comparison MCI vs. CN, no signiﬁcant

DEGs were found using SAM; thus, the genes to be

included in the integrated classiﬁcation were obtained

by selecting the top 300 genes with the highest vari-

ance from the expression matrix of the two classes

samples.

3 EXPERIMENTS

3.1 Evaluation Procedure

For classiﬁcation, we adopted an SVM with linear

kernel. The evaluation consists of 10 iterations of

5-fold cross-validation, using stratiﬁed partitions of

the data into train and test subsets. At each itera-

tion, training folds are z-scored, and their mean and

variance are used to z-score the test set accordingly.

The performance results have been computed as av-

erage over the iterations of the well-known metrics

summarized in Table 2. These are deﬁned in terms

of the number of true positives (TP), true negatives

(TN), false positives (FP), and false negatives (FN).

Here, the ﬁrst class in each task (e.g., AD in AD vs.

CN) is assumed as the positive class. All the metrics

assume values in [0,1], except MCC that ranges in

[-1,1]; higher values indicate better performance for

all the metrics.

3.2 Performance Results

Table 3 reports performance results obtained for each

binary classiﬁcation problem by adopting only imag-

ing features (MRI), only omics features (GE), or both

(MRI+GE). Here, it can be observed that extremely

good performance is achieved for the AD vs. CN

task, with MRI+GE features leading to the best results

against MRI and GE taken separately. Indeed, even

though imaging features lead to better performance

than omics, their combination leads to increased per-

formance in all the metrics.

The remaining two binary classiﬁcation tasks are

notoriously hard, and thus lower performance is

achieved. For the AD vs. MCI task, omics data

alone leads to slightly higher performance values than

imaging data alone, but their combination still leads to

increased performance. Instead, for the MCI vs. CN

task, omics data leads to such a poor performance that

its inﬂuence in the combined features leads to results

worse than using imaging features alone.

Finally, it can be observed that class unbalanc-

ing in the two binary tasks that include AD patients

(positive minority class, including a low number of

samples) leads to a much higher recognition rate for

the negative majority classes, experiencing speciﬁcity

much higher than sensitivity. On the other side, F-

measure, AUC, MCC, and BA conﬁrm to be metrics

less dependent on class unbalancing and well support

the exposed performance analysis.

The main observations arising from the analysis

of the results are 1) in most cases, the imaging fea-

tures perform much better than the omics features; 2)

for classifying AD against CN or MCI patients, the

combination of omics and imaging features leads to

better results than the same features taken separately;

3) the MCI vs. CN task still needs to be investigated,

as none of the considered sets of features leads to ac-

ceptable performance results.

3.3 Comparison with the

State-of-the-Art

In Table 4, we report classiﬁcation performance re-

sults on ADNI data published in recent literature, also

specifying the cardinality of the subsets of samples

considered (column ‘# Samples’) and the type of fea-

tures adopted (column ‘Feats.’). Even though all the

reported results have been obtained using different

subsets of ADNI data and varying evaluation pro-

tocols, the Table intends to provide a rough perfor-

mance comparison of the achieved results. Our best

results from Table 3 are also reported to make a more

immediate comparison.

(Cheng and Liu, 2017) constructed multi-level

CNNs to gradually learn and combine multi-modal

features for AD classiﬁcation extracted by MRI and

PET images. First, two deep 3D-CNNs are con-

structed to transform the whole brain information into

compact high-level features for each modality. Then,

differently from conventional combination methods

that average the class probabilistic scores, a 2D CNN

is learned to combine the multi-modal features and

make the ﬁnal classiﬁcation.

In (Aderghal et al., 2017), a CNN is trained on

features extracted from the hippocampal region from

MRIs, using data augmentation strategies to obtain

the needed large volumes of data and data balanc-

ing strategies to handle unbalanced classes. Later on,

the same group (Aderghal et al., 2018) proposed a

Classifying Alzheimer’s Disease using MRIs and Transcriptomic Data

Table 2: Performance measures adopted in the experiments.

Acron. Name Formula Description

Acc Accuracy

TP + TN

TP + FN+ FP + TN

% of correctly classiﬁed samples

Spec Speciﬁcity

TN + FP

% of negative samples correctly

identiﬁed

Sens

Sensitivity

(or Recall

or TPR)

TP + FN

% of positive samples correctly

classiﬁed

Prec Precision

TP + FP

% of positive samples correctly

classiﬁed, considering the set of all

the samples classiﬁed as positive

F-measure

2 · Prec · Sens

Prec + Sens

Weighted compromise between

Sens and Prec

Gm G-mean

Sens · Spec

Geometric mean of the accuracy of

both classes

AUC

Area Under

the ROC

Curve

Sens(x)dx, x = 1−Spec

Uses the ROC curve to exhibit the

trade-off between the classiﬁer’s TP

and FP rates

MCC

Matthews

Correlation

Coefﬁcient

(Matthews,

1975)

TP · TN − FP · FN

(TP + FP)(TP + FN)(TN + FP)(TN + FN)

Correlation coefﬁcient between ob-

served and predicted binary clas-

siﬁcations. Useful for unbalanced

classes

Balanced

Accuracy

Sens + Spec

Compromise between Spec and

Sens. Useful for unbalanced classes

Table 3: Average performance results for the three binary classiﬁcation problems using only imaging features (MRI), only

omics features (GE), or both (MRI+GE). In boldface the best values for each metric and each classiﬁcation problem.

Features Acc Sens Spec Prec F1 Gm AUC MCC BA

AD vs. CN

MRI 0.927 0.632 0.977 0.828 0.707 0.780 0.926 0.680 0.804

GE 0.860 0.479 0.924 0.538 0.492 0.655 0.847 0.422 0.701

MRI+GE 0.946 0.722 0.983 0.889 0.787 0.839 0.955 0.768 0.853

AD vs. MCI

MRI 0.871 0.244 0.932 0.258 0.245 0.444 0.722 0.178 0.588

GE 0.878 0.342 0.931 0.330 0.330 0.550 0.774 0.267 0.637

MRI+GE 0.915 0.394 0.966 0.546 0.448 0.606 0.869 0.415 0.680

MCI vs. CN

MRI 0.636 0.732 0.471 0.704 0.717 0.585 0.651 0.207 0.601

GE 0.524 0.601 0.394 0.629 0.614 0.484 0.499 -0.005 0.497

MRI+GE 0.562 0.632 0.443 0.661 0.645 0.526 0.555 0.073 0.537

method that combines the MRI and DTI (Diffusion

Tensor Imaging) modalities. Due to the scarce avail-

ability of DTIs, they adopted cross-modal transfer

learning from MRIs to DTIs and combined the classi-

ﬁcation results of multiple CNNs by a majority vote.

(Tong et al., 2017) presented a multi-modality

classiﬁcation framework to exploit the complemen-

tarity in the multi-modal data. They ﬁrst compute

pairwise similarity for each modality individually us-

ing features from regional MRI volumes, voxel-based

FDG-PET signal intensities, CSF biomarker mea-

sures, and APOE4 genetic information. Then, they

combine the similarities in a nonlinear graph fusion

process, which generates a uniﬁed graph for ﬁnal clas-

siﬁcation.

In (B

ackstr

om et al., 2018), a 3D CNN is pro-

posed, named 3DConvNet, for AD vs. CN classi-

ﬁcation. It consists of ﬁve convolutional layers for

feature extraction from MRIs, followed by three fully

connected layers for classiﬁcation.

BIOIMAGING 2022 - 9th International Conference on Bioimaging

Table 4: Performance comparisons of recent classiﬁcation methods on the ADNI dataset. In boldface the best values for each

metric and each classiﬁcation problem.

Ref. # Samples Feats. Acc Sens Spec Prec F

AUC BA

AD vs. CN

(Aderghal et al., 2017) 188, 228 MRI 0.828 0.796 0.859 - - - 0.828

(Cheng and Liu, 2017) 93, 100 MRI,

PET

0.896 0.871 0.920 - - 0.945 0.896

(Tong et al., 2017) 37, 35 APOE,

CSF,

MRI,

PET

0.918 0.889 0.947 - - 0.983 0.918

(Aderghal et al., 2018) 236, 285 MRI,

DTI

0.925 0.947 0.904 - - - 0.925

ackstr

om et al., 2018) 199, 141 MRI 0.901 0.933 0.868 - - - 0.900

(Li and Liu, 2018) 199, 229 MRI 0.897 0.879 0.908 - - 0.924 0.894

(Senanayake et al., 2018) 161, 161 MRI,

Cog.

tests

0.760 - - - - - -

(Shi et al., 2018) 51, 52 MRI,

PET

0.971 0.959 0.985 - - - 0.972

(Gupta et al., 2019) 38, 38 APOE,

CSF,

MRI,

PET

0.984 1.000 0.965 0.979 0.984 - 0.983

(Bae et al., 2020) 195, 195 MRI 0.890 0.880 0.910 - - 0.940 0.895

(Lee and Lee, 2020) 63, 136 GE - - - - - 0.665 -

Our best results 42, 250 MRI,

0.946 0.722 0.983 0.889 0.787 0.955 0.853

AD vs. MCI

(Aderghal et al., 2017) 188, 199 MRI 0.660 0.737 0.587 - - - 0.662

(Aderghal et al., 2018) 236, 503 MRI,

DTI

0.850 0.937 0.791 - - - 0.864

(Senanayake et al., 2018) 161, 193 MRI,

Cog.

tests

0.760 - - - - - -

Our best results 42, 428 MRI,

0.915 0.394 0.966 0.546 0.448 0.869 0.680

MCI vs. CN

(Aderghal et al., 2017) 199, 228 MRI 0.625 0.600 0.640 - - - 0.620

(Tong et al., 2017) 75, 35 APOE,

CSF,

MRI,

PET

0.795 0.851 0.671 - - 0.812 0.761

(Aderghal et al., 2018) 503, 285 MRI,

DTI

0.800 0.928 0.730 - - - 0.829

(Li and Liu, 2018) 403, 229 MRI 0.738 0.866 0.515 - - 0.775 0.802

(Senanayake et al., 2018) 193, 161 MRI,

Cog.

tests

0.750 - - - - - -

(Shi et al., 2018) 99, 52 MRI,

PET

0.872 0.979 0.670 - - - 0.825

Our best results 428, 250 MRI 0.636 0.732 0.471 0.704 0.717 0.651 0.601

Classifying Alzheimer’s Disease using MRIs and Transcriptomic Data

(Li and Liu, 2018) proposed a classiﬁcation

method based on multiple cluster dense convolutional

neural networks (DenseNets) to learn features from

MRIs. Each whole-brain image is ﬁrst partitioned

into different local regions, and a ﬁxed number of

3D patches is extracted from each region. Then, the

patches from each region are grouped into different

clusters with k-means clustering. A DenseNet is con-

structed to learn the patch features for each cluster

and the features learned from the discriminative clus-

ters of each region are ensembled for classiﬁcation.

Finally, the classiﬁcation results from different local

regions are combined to enhance the ﬁnal image clas-

siﬁcation.

(Senanayake et al., 2018) used 3D MR volumes

and neuropsychological measure-based (NM) feature

vectors. For combining these two data sources, hav-

ing very different dimensions (35 NM features against

more than ten million features from 3D MR volumes),

they proposed a deep learning-based pipeline that re-

duces the dimension of the MRI features to a dimen-

sion comparable with that of NM, and used the feature

vector merging the two sets of features.

(Shi et al., 2018) proposed a multi-modal algo-

rithm based on a stacked deep polynomial network

(MM-SDPN). Two SDPNs are ﬁrst used to learn

high-level features from MRIs and PETs separately,

which are then fed to another SDPN to fuse multi-

modal neuroimaging information to contain the in-

trinsic properties of both modalities and their corre-

lation.

(Gupta et al., 2019) proposed a machine learning-

based framework, based on SVM and feature selec-

tion, to discriminate the various stages of ADNI pa-

tients using a combination of FDG-PET, structural

MRI, CSF protein levels, and APOE genotype. Here,

the MCI group of patients is subdivided into MCIc

(MCI converted, i.e., MCI patients that converted to

AD within 24 months) and MCIs (MCI stable, i.e.,

that did not convert to AD within 24 months); there-

fore, their interesting conclusions on binary prob-

lems involving MCI patients cannot be compared with

ours.

(Bae et al., 2020) developed a CNN-based algo-

rithm to classify AD patients and CN controls using

coronal slices of T1-weighted MRI images that cover

the medial temporal lobe. They tested it on two inde-

pendent populations, including ADNI patients.

(Lee and Lee, 2020) classiﬁed AD vs. CN using

blood gene expression data. They tested ﬁve feature

selection methods and ﬁve classiﬁers. The best AUC

in the internal evaluation on the ADNI dataset was ob-

tained using DEGs extracted using SAM without fea-

ture selection and using a deep neural network classi-

ﬁer.

Table 4 shows that our results using MRI and GE

features are competitive with those achieved by state-

of-the-art methods for the AD vs. CN and AD vs.

MCI classiﬁcation tasks. It is interesting to observe

that the highest performance results are reported for

methods (e.g., (Tong et al., 2017; Gupta et al., 2019))

that take into account not only MRI and PET features

but also CSF. However, the extraction of such data

requires a quite invasive intervention, preventing us

from adopting them in our multi-modal setting.

Moreover, it should be explicitly observed that the

neuropsychological measures adopted as features by

some methods (e.g., (Senanayake et al., 2018)) are

generally considered by medical doctors to diagnose

the disease state of each patient. Thus, their use as

features for classiﬁcation appears to strongly and pos-

itively bias the results. This is shown in Table 5,

where we report extremely high performance results

achieved with our classiﬁcation procedure using as

features only three cognitive tests (CDRSB, ADAS11,

and MMSE) on the selected subset of samples. Simi-

lar results are achieved on the whole set of data from

ADNIMERGE, as shown in Table 6.

4 CONCLUSIONS

In this paper, we propose a method for classifying the

various stages of Alzheimer’s disease, which relies on

data acquired by non-invasive techniques and that are

compatible with the limitations imposed by pandemic

situations. The multi-modal data consist of omics and

imaging features extracted by gene expression values

from blood samples and MRIs, respectively. We show

how a suitable integration of omics and imaging data,

using well-known machine learning techniques, can

lead to better results than any of them taken separately

for the classiﬁcation of AD against CN or MCI pa-

tients. Moreover, the achieved performance appears

competitive with the state-of-the-art. However, when

discriminating MCI and CN patients, none of the con-

sidered sets of features leads to acceptable perfor-

mance results. This classiﬁcation task, well known

to be more challenging than the other two, needs to

be further investigated.

ACKNOWLEDGEMENTS

This work has been partially funded by the BiBi-

Net project (H35F21000430002) within POR-Lazio

FESR 2014-2020. It was carried out also within the

BIOIMAGING 2022 - 9th International Conference on Bioimaging

Table 5: Average performance results for the three binary classiﬁcation problems using as features only three cognitive tests

(CDRSB, ADAS11, and MMSE) on the considered ADNI subset (42 AD, 427 MCI, 250 CN).

Task Acc Sens Spec Prec F1 Gm AUC MCC BA

AD vs. CN 0.990 0.976 0.992 0.958 0.965 0.984 0.996 0.961 0.984

AD vs. MCI 0.835 0.976 0.821 0.355 0.519 0.895 0.903 0.529 0.899

MCI vs. CN 0.867 0.926 0.768 0.873 0.898 0.841 0.931 0.716 0.847

Table 6: Average performance results for the three binary classiﬁcation problems using as features only three cognitive tests

(CDRSB, ADAS11, and MMSE) on the whole ADNIMERGE dataset (397 AD, 1055 MCI, 519 CN).

Task Acc Sens Spec Prec F1 Gm AUC MCC BA

AD vs. CN 0.994 0.986 1.000 1.000 0.993 0.993 1.000 0.988 0.993

AD vs. MCI 0.930 0.813 0.975 0.924 0.864 0.890 0.978 0.821 0.894

MCI vs. CN 0.922 0.915 0.937 0.968 0.940 0.926 0.979 0.832 0.926

activities of the authors as members of the ICAR-

CNR INdAM Research Unit and partially supported

by the INdAM research project “Computational In-

telligence methods for Digital Health”. The work

of Mario R. Guarracino was conducted within the

framework of the Basic Research Program at the Na-

tional Research University Higher School of Eco-

nomics (HSE). Mario Manzo thanks Prof. Alfredo

Petrosino for the guidance and supervision during the

years of working together.

Data collection and sharing for this project was

funded by the Alzheimer’s Disease Neuroimaging

Initiative (ADNI) (National Institutes of Health Grant

U01 AG024904) and DOD ADNI (Department of De-

fense award number W81XWH-12-2-0012). ADNI

is funded by the National Institute on Aging, the

National Institute of Biomedical Imaging and Bio-

engineering, and through generous contributions from

the following: AbbVie, Alzheimer’s Association;

Alzheimer’s Drug Discovery Foundation; Araclon

Biotech; BioClinica, Inc.; Biogen; Bristol-Myers

Squibb Company; CereSpir, Inc.; Cogstate; Eisai

Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Com-

pany; EuroImmun; F. Hoffmann-La Roche Ltd and

its afﬁliated company Genentech, Inc.; Fujirebio;

GE Healthcare; IXICO Ltd.; Janssen Alzheimer Im-

munotherapy Research & Development, LLC.; John-

son & Johnson Pharmaceutical Research & Devel-

opment LLC.; Lumosity; Lundbeck; Merck & Co.,

Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Re-

search; Neurotrack Technologies; Novartis Pharma-

ceuticals Corporation; Pﬁzer Inc.; Piramal Imag-

ing; Servier; Takeda Pharmaceutical Company; and

Transition Therapeutics. The Canadian Institutes of

Health Research is providing funds to support ADNI

clinical sites in Canada. Private sector contributions

are facilitated by the Foundation for the National In-

stitutes of Health (www.fnih.org). The grantee orga-

nization is the Northern California Institute for Re-

search and Education, and the study is coordinated by

the Alzheimer’s Therapeutic Research Institute at the

University of Southern California. ADNI data are dis-

seminated by the Laboratory for Neuro Imaging at the

University of Southern California.

REFERENCES

Aderghal, K., Boissenin, M., Benois-Pineau, J., Catheline,

G., and Afdel, K. (2017). Classiﬁcation of sMRI

for AD diagnosis with convolutional neuronal net-

works: A pilot 2-D+ε study on ADNI. In Amsaleg,

L. et al., editors, MultiMedia Modeling - 23rd Inter-

national Conference, MMM 2017, Reykjavik, Iceland,

January 4-6, 2017, Proceedings, Part I, volume 10132

of Lecture Notes in Computer Science, pages 690–

701. Springer.

Aderghal, K., Khvostikov, A., Krylov, A., Benois-Pineau,

J., Afdel, K., and Catheline, G. (2018). Classiﬁcation

of Alzheimer disease on imaging modalities with deep

CNNs using cross-modal transfer learning. In 2018

IEEE 31st International Symposium on Computer-

Based Medical Systems (CBMS), pages 345–350.

Antonelli, L., Guarracino, M. R., Maddalena, L., and San-

giovanni, M. (2019). Integrating imaging and omics

data: A review. Biomedical Signal Processing and

Control, 52:264–280.

Ashburner, J. (2007). A fast diffeomorphic image registra-

tion algorithm. NeuroImage, 38(1):95–113.

Ashburner, J. and Friston, K. J. (2005). Uniﬁed segmenta-

tion. NeuroImage, 26(3):839–851.

Bae, J., Lee, S., Jung, W., Park, S., Kim, W., Oh, H., Han,

J., Kim, G., Kim, J., Kim, J., and Kim, K. (2020).

Identiﬁcation of Alzheimer’s disease using a convo-

lutional neural network model based on T1-weighted

magnetic resonance imaging. Scientiﬁc reports, 10(1).

ackstr

om, K., Nazari, M., Gu, I. Y.-H., and Jakola, A. S.

(2018). An efﬁcient 3D deep convolutional net-

work for Alzheimer’s disease diagnosis using MR im-

ages. In 2018 IEEE 15th International Symposium on

Biomedical Imaging (ISBI 2018), pages 149–153.

Birkenbihl, C., Salimi, Y., Domingo-Fernandez, D., on be-

Classifying Alzheimer’s Disease using MRIs and Transcriptomic Data

half of the AddNeuroMed consortium, S. L., Fr

ohlich,

H., Hofmann-Apitius, M., the Japanese Alzheimer’s

Disease Neuroimaging Initiative, and the Alzheimer’s

Disease Neuroimaging Initiative (2020). Evaluating

the Alzheimer’s disease data landscape. Alzheimer’s

& Dementia: Translational Research & Clinical In-

terventions, 6(1):e12102.

Birkenbihl, C., Westwood, S., Shi, L., Nevado-Holgado,

A., Westman, E., Lovestone, S., Consortium, A., and

Hofmann-Apitius, M. (2021). ANMerge: A compre-

hensive and accessible Alzheimer’s disease patient-

level dataset. J Alzheimers Dis., 79:423–431.

Cheng, D. and Liu, M. (2017). CNNs based multi-modality

classiﬁcation for AD diagnosis. In 2017 10th In-

ternational Congress on Image and Signal Process-

ing, BioMedical Engineering and Informatics (CISP-

BMEI), pages 1–5.

Gorgolewski, K., Auer, T., Calhoun, V., Craddock, R.,

Das, S., Duff, E., Flandin, G., Ghosh, S., Glatard, T.,

Halchenko, Y., Handwerker, D., Hanke, M., Keator,

D., Li, X., Michael, Z., Maumet, C., Nichols, B.,

Nichols, T., Pellman, J., Poline, J., Rokem, A., Schae-

fer, G., Sochat, V., Triplett, W., Turner, J., Varoquaux,

G., and Poldrack, R. (2016). The brain imaging data

structure, a format for organizing and describing out-

puts of neuroimaging experiments. Scientiﬁc data, 3.

reserved.

Gupta, Y., Lama, R. K., Kwon, G.-R., et al. (2019). Predic-

tion and classiﬁcation of Alzheimer’s disease based

on combined features from apolipoprotein-e geno-

type, cerebrospinal ﬂuid, MR, and FDG-PET imag-

ing biomarkers. Frontiers in Computational Neuro-

science, 13:72.

Joliot, M., Jobard, G., Naveau, M., Delcroix, N., Petit, L.,

Zago, L., Crivello, F., Mellet, E., Mazoyer, B., and

Tzourio-Mazoyer, N. (2015). AICHA: An atlas of in-

trinsic connectivity of homotopic areas. Journal of

Neuroscience Methods, 254:46–59.

Lawrence, E., Vegvari, C., Ower, A., Hadjichrysanthou, C.,

De Wolf, F., and RM, A. (2017). A systematic review

of longitudinal studies which measure Alzheimer’s

disease biomarkers. J Alzheimers Dis., 59(4):1359–

1379.

Lee, T. and Lee, H. (2020). Prediction of Alzheimer’s

disease using blood gene expression data. Sci Rep,

10(1):3485.

Li, F. and Liu, M. (2018). Alzheimer’s disease diagno-

sis based on multiple cluster dense convolutional net-

works. Computerized Medical Imaging and Graphics,

70:101–110.

Li, X., Wang, H., Long, J., et al. (2018). Systematic analysis

and biomarker study for Alzheimer’s disease. Sci Rep,

8:17394.

Lovestone, S., Francis, P., Kloszewska, I., Mecocci, P., Sim-

mons, A., Soininen, H., Spenger, C., Tsolaki, M., Vel-

las, B., Wahlund, L., Ward, M., and Consortium, A.

(2009). AddNeuroMed–the European collaboration

for the discovery of novel biomarkers for Alzheimer’s

disease. Ann N Y Acad Sci, pages 36–46.

Maddalena, L., Granata, I., Manipur, I., Manzo, M., and

Guarracino, M. (2020). Glioma grade classiﬁcation

via omics imaging. In Proceedings of the 13th Inter-

national Joint Conference on Biomedical Engineering

Systems and Technologies - Volume 2: BIOIMAGING,

pages 82–92. INSTICC, SciTePress.

Maddalena, L., Granata, I., Manipur, I., Manzo, M., and

Guarracino, M. R. (2021). A framework based on

metabolic networks and biomedical images data to

discriminate glioma grades. In Ye, X., Soares, F.,

De Maria, E., G

omez Vilda, P., Cabitza, F., Fred,

A., and Gamboa, H., editors, Biomedical Engineer-

ing Systems and Technologies, pages 165–189, Cham.

Springer International Publishing.

Mart

ı-Juan, G., Sanroma-Guell, G., and Piella, G. (2020).

A survey on machine and statistical learning for longi-

tudinal analysis of neuroimaging data in Alzheimer’s

disease. Comput. Methods Programs Biomed.,

189:105348.

Matthews, B. (1975). Comparison of the predicted and

observed secondary structure of T4 phage lysozyme.

Biochimica et Biophysica Acta (BBA) - Protein Struc-

ture, 405(2):442–451.

Mueller, S., Weiner, M., Thal, L., Petersen, R., Jack, C.,

Jagust, W., Trojanowski, J., Toga, A., and Beck-

ett, L. (2005). Ways toward an early diagnosis in

Alzheimer’s disease: the Alzheimer’s Disease Neu-

roimaging initiative (ADNI). J Alzheimers Dement.,

1(1):55–66.

Nho, K., ADNI, et al. (2016). Integration of bioinformatics

and imaging informatics for identifying rare PSEN1

variants in Alzheimer’s disease. BMC Medical Ge-

nomics, 9(Suppl 1).

Peng, J., An, L., Zhu, X., Jin, Y., and Shen, D. (2016).

Structured sparse kernel learning for imaging genet-

ics based Alzheimer’s disease diagnosis. In Inter-

national Conference on Medical Image Computing

and Computer-Assisted Intervention, pages 70–78.

Springer.

Routier, A., Burgos, N., D

ıaz, M., Bacci, M., Bottani, S., El-

Rifai, O., Fontanella, S., Gori, P., Guillon, J., Guyot,

A., Hassanaly, R., Jacquemont, T., Lu, P., Marcoux,

A., Moreau, T., Samper-Gonz

alez, J., Teichmann, M.,

Thibeau-Sutre, E., Vaillant, G., Wen, J., Wild, A.,

Habert, M.-O., Durrleman, S., and Colliot, O. (2021).

Clinica: An open-source software platform for re-

producible clinical neuroscience studies. Frontiers in

Neuroinformatics, 15:39.

Samper-Gonz

alez, J., Burgos, N., Bottani, S., Fontanella,

S., Lu, P., Marcoux, A., Routier, A., Guillon, J., Bacci,

M., Wen, J., Bertrand, A., Bertin, H., Habert, M. O.,

Durrleman, S., Evgeniou, T., and Colliot, O. (2018).

Reproducible evaluation of classiﬁcation methods in

Alzheimer’s disease: Framework and application to

MRI and PET data. NeuroImage, 183:504–521.

Senanayake, U., Sowmya, A., and Dawes, L. (2018). Deep

fusion pipeline for mild cognitive impairment diagno-

sis. In 2018 IEEE 15th International Symposium on

Biomedical Imaging (ISBI 2018), pages 1394–1997.

Shi, J., Zheng, X., Li, Y., Zhang, Q., and Ying, S. (2018).

Multimodal neuroimaging feature learning with mul-

BIOIMAGING 2022 - 9th International Conference on Bioimaging

timodal stacked deep polynomial networks for di-

agnosis of Alzheimer’s disease. IEEE Journal of

Biomedical and Health Informatics, 22(1):173–183.

Tong, T., Gray, K., Gao, Q., Chen, L., and Rueckert, D.

(2017). Multi-modal classiﬁcation of Alzheimer’s dis-

ease using nonlinear graph fusion. Pattern Recogni-

tion, 63:171–181.

Tusher, V. G., Tibshirani, R., and Chu, G. (2001). Signif-

icance analysis of microarrays applied to the ioniz-

ing radiation response. Proceedings of the National

Academy of Sciences, 98(9):5116–5121.

Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D.,

Crivello, F., Etard, O., Delcroix, N., Mazoyer, B.,

and Joliot, M. (2002). Automated anatomical labeling

of activations in SPM using a macroscopic anatomi-

cal parcellation of the MNI MRI single-subject brain.

NeuroImage, 15(1):273–289.

Vapnik, V. (1995). The Nature of Statistical Learning The-

ory. Springer-Verlag.

Voyle, N., Keohane, A., Newhouse, S., Lunnon, K., John-

ston, C., Soininen, H., Kloszewska, I., Mecocci, P.,

Tsolaki, M., Vellas, B., Lovestone, S., Hodges, A.,

Kiddle, S., and Dobson, R. (2016). A pathway based

classiﬁcation method for analyzing gene expression

for Alzheimer’s disease diagnosis. J Alzheimers Dis,

49(3):659–69.

Wen, J., Thibeau-Sutre, E., Diaz-Melo, M., Samper-

Gonz

alez, J., Routier, A., Bottani, S., Dormont, D.,

Durrleman, S., Burgos, N., and Colliot, O. (2020).

Convolutional neural networks for classiﬁcation of

Alzheimer’s disease: Overview and reproducible eval-

uation. Medical Image Anal., 63:101694.

WHO (2019). Risk reduction of cognitive decline and de-

mentia. World Health Organization Guidelines.

Classifying Alzheimer’s Disease using MRIs and Transcriptomic Data