RESEARCH ARTICLE

Diagnosis of Major Depressive Disorder

Using Machine Learning Based on
Multisequence MRI Neuroimaging Features
Qinghe Li, MD,1,2 Fanghui Dong, MD,3 Qun Gai, MD,3 Kaili Che, MD,3 Heng Ma, MD,3
Feng Zhao, PhD,4 Tongpeng Chu, MSc,3* Ning Mao, MD, PhD,3* and
Peiyuan Wang, MD1*

Background: Previous studies have found qualitative structural and functional brain changes in major depressive disorder
(MDD) patients. However, most studies ignored the complementarity of multisequence MRI neuroimaging features and
cannot determine accurate biomarkers.
Purpose: To evaluate machine-learning models combined with multisequence MRI neuroimaging features to diagnose
patients with MDD.
Study Type: Prospective.
Subjects: A training cohort including 111 patients and 90 healthy controls (HCs) and a test cohort including 28 patients
and 22 HCs.
Field Strength/Sequence: A 3.0 T/T1-weighted imaging, resting-state functional MRI with echo-planar sequence, and
single-shot echo-planar diffusion tensor imaging.
Assessment: Recruitment and integration were used to reflect the dynamic changes of functional networks, while gray
matter volume and fractional anisotropy were used to reflect the changes in the morphological and anatomical network.
We then fused features with significant differences in functional, morphological, and anatomical networks to evaluate a
random forest (RF) classifier to diagnose patients with MDD. Furthermore, a support vector machine (SVM) classifier was
used to verify the stability of neuroimaging features. Linear regression analyses were conducted to investigate the relation-
ships among multisequence neuroimaging features and the suicide risk of patients.
Statistical Tests: The comparison of functional network attributes between patients and controls by two-sample t-test.
Network-based statistical analysis was used to identify structural and anatomical connectivity changes between MDD and
HCs. The performance of the model was evaluated by receiver operating characteristic (ROC) curves.
Results: The performance of the RF model integrating multisequence neuroimaging features in the diagnosis of depres-
sion was significantly improved, with an AUC of 93.6%. In addition, we found that multisequence neuroimaging features
could accurately predict suicide risk in patients with MDD (r = 0.691).
Data Conclusion: The RF model fusing functional, morphological, and anatomical network features performed well in diag-
nosing patients with MDD and provided important insights into the pathological mechanisms of MDD.
Evidence Level: 1.
Technical Efficacy: Stage 2.
J. MAGN. RESON. IMAGING 2023;58:1420–1430.

View this article online at wileyonlinelibrary.com. DOI: 10.1002/jmri.28650

Received Nov 26, 2022, Accepted for publication Feb 4, 2023.

*Address reprint requests to: P.W., Department of Radiology, Yantai Affiliated Hospital of Binzhou Medical University, Yantai, Shandong, People’s Republic of
China, 264100. E-mail: wangpeiyuan1640@163.com, N.M., Department of Radiology, Yantai Yuhuangding Hospital, Qingdao University, Yantai, Shandong, Peo-
ple’s Republic of China, 264000. E-mail: maoning@pku.edu.cn, or T.C., Department of Radiology, Yantai Yuhuangding Hospital, Qingdao University, Yantai,
Shandong, People’s Republic of China, 264000. E-mail: chutongpeng@163.com
Qinghe Li and Fanghui Dong are co-first authors on the paper.
Contract grant sponsor: National Natural Science Foundation of China; Contract grant number: 82001775, 62176140, 61773244; Contract grant sponsor: Natural
Science Foundation of Shandong Province; Contract grant number: ZR2021MH120; Contract grant sponsor: “Taishan Scholar”; Contract grant number:
tsqn202103197.

From the 1Department of Radiology, Yantai Affiliated Hospital of Binzhou Medical University, Yantai, Shandong, People’s Republic of China; 2School of
Medical Imaging, Binzhou Medical University, Yantai, Shandong, People’s Republic of China; 3Department of Radiology, Yantai Yuhuangding Hospital,
Qingdao University, Yantai, Shandong, People’s Republic of China; and 4School of Compute Science and Technology, Shandong Technology and Business
University, Yantai, Shandong, People’s Republic of China
Additional supporting information may be found in the online version of this article

1420 © 2023 International Society for Magnetic Resonance in Medicine.


M ajor depressive disorder (MDD) is a neuro-
heterogeneous disease with a high recurrence rate and
high mortality, which mainly manifests as anxiety, depression,
and persistent low mood.1 It not only affects the individual’s
physical and mental health and social functions but also
increases social burden.2 According to the World Health
Organization (WHO), MDD will become the world’s largest
disease burden by 2030.3 As one of the most common mental
disorders, the onset of MDD is often hidden, and there is
currently no accurate biomarker.
As a noninvasive method, MRI provides important
insights into the neuropathology of MDD. Structural MRI
studies that have previously quantified morphological changes
in the gray matter indicated that MDD-related atrophy
mainly occurs in the occipital lobe and hippocampus.4,5
Diffusion-weighted imaging has been used to describe the
movement of water molecules in white matter. Compared
with HCs, the integrity of the thalamus and cingulate gyrus
fibers in patients with MDD is decreased.6 Resting-state func-
tional MRI (rs-fMRI), which measures the spontaneous activ-
ity of blood oxygen levels in different brain regions, indicates
that the resting-state functional connection of the amygdala
in patients with MDD is impaired.7 Although each MRI
sequence provides unique information related to MDD, when
evaluated independently, the study of a single sequence may
provide incomplete information to diagnose MDD. Mean-
while, when exploring and identifying potential abnormalities
occurring in MDD, an increasing number of studies have
supported the complementarity of brain information in differ-
ent sequences.8,9
Multisequence brain image analysis has been used in
clinical research on MDD and has research value.9,10 For
example, Sun et al used the combination of gray matter vol-
ume calculated from structural MRI, and amplitude low-
frequency fluctuations (ALFF) from functional MRI as input
features to diagnose MDD from HCs.10 Yang et al merged
regional volumetric measures with regional functional connec-
tions for MDD classification using a linear support vector
machine (SVM).9 Han et al used a combination of volumes
extracted from structural MRI and functional connectivity
calculated from functional MRI as features for MDD classifi-
cation.11 Although these multisequence studies have
improved the performance of MDD diagnosis, these measure-
ments assume that the neural communication of the whole
brain is static, ignoring the dynamic characteristics of the
brain.
The human brain can be thought of as a complex
dynamic brain network, that can spontaneously form com-
munity structures and be reconstructed over time.12 Investi-
gating temporal fluctuations in MDD brain network
topology might advance our understanding of how dynamic
interactions of different network components underpin clini-
cal symptoms in patients. Advances in machine learning
November 2023
15222586, 2023, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jmri.28650 by Manager Information Resources Bond University Library, Wiley Online Library on [14/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Li et al.: Diagnosis of Major Depressive Disorder
techniques have shown great potential for the early diagnosis
of mental illnesses.9 However, there are many variations in
the MDD diagnosis results between machine learning
methods combined with multisequence neuroimaging fea-
tures, which warrant further exploration.10,13
The first goal of this study was to calculate the two
attributes of brain regions based on the multilayer network of
dynamic functional connections and fuse morphological and
anatomical network features to diagnose patients with MDD.
The second aim was to explore the relationship between mul-
tisequence imaging biomarkers and suicide risk in patients
with MDD and to examine whether these biomarkers can be
used as baseline predictors of clinical suicide risk.
Materials and Methods
Participants
This study was approved by the Research Ethics Committee of our
hospital. Before the start of the experiment, all participants were
informed of the purpose, procedure, and related contraindications of
the experiment. Informed consent was obtained from all subjects. In
this prospective study, we recruited 163 right-handed MDD patients
and 135 well-matched HCs. The patients were recruited from the
psychological clinic of our hospital and met the diagnostic criteria
and Diagnostic Statistical Manual of Mental Disorders Fifth Edition.
HCs were recruited from the local community that matched the
MDD group in terms of age, education level, and body mass index
(BMI). Details of exclusion and inclusion criteria for MDD and
HCs are included in the Supplementary Material S1.
MRI Protocol
In this study, all subjects were scanned with a 3.0 T MR
(GE 750W, GE Healthcare, USA) equipped with an eight-channel
head coil. The MRI protocol included: rs-fMRI with gradient echo
sequence, high-resolution 1
1
1 T1-weight sequence, and
single-shot echo-planar DTI sequence. Image parameters are shown
in Supplementary Material S1.
Image Preprocessing and Network
Construction
Data Preprocessing
Figure 1 shows the workflow of this study. Functional imag-
ing data were preprocessed using Data Processing Assistant
for Resting-State fMRI (DPARSF v5.2; http://www.restfmri.
net/forum/dparsf).14 The data processing steps included Slice
timing and head-motion corrections, normalization, filtering,
and multiple linear regression analysis. Detailed information
is included in the Supplementary Material S1.
The DTI data were preprocessed using pipeline for ana-
lyzing brain diffusion images (PANDA, http://www.nitrc.org/
projects/panda/). First, converte the DTI and 3D-T1 data
from DICOM format to the 4D NIfTI format. Next,
head movement and eddy current correction is performed to
correct image distortion. The DTI data are co-registered
to the high-resolution 3D-T1 images, subsequently, the
1421
 15222586, 2023, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jmri.28650 by Manager Information Resources Bond University Library, Wiley Online Library on [14/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Magnetic Resonance Imaging

FIGURE 1: Workflow of the analysis. (a) Multimodel brain network modeling. Constructing morphological, anatomical, and functional
networks separately. (b) Feature extraction and selection of brain network. (c) Prediction. Machine learning model is used to
diagnose MDD and predict the suicide risk in patients with MDD. MDD = major depressive disorder; LASSO = least absolute
shrinkage and selection operator; RF = random forest; SVM = support vector machine.

co-registered images were normalized into the Montreal Neu-
rological Institute (MNI) space.
Structural image data were preprocessed using the Com-
putational Anatomy Toolbox (CAT12, http://www.neuro.
uni-jena.de/cat/). Each image was checked by an experienced
neuroradiologist for artifacts and manually reoriented to
adjust the origin of an image at the anterior commissure.
Next, the structural T1-weighted images were normalized to
MNI space and resampled to 3
3
3 mm. Then, all the
normalized images were segmented into gray matter (GM),
white matter, and cerebrospinal fluid.
Functional Network
For each individual, we used the sliding window
approach15 to construct dynamic brain connectomes in the
Dynamic BC toolbox (v2.2; http://restfmri.net/forum/
DynamicBC). Specifically, the whole brain was divided
into 90 regions according to the anatomical automatic
labeling (AAL) template,16 and the average fMRI time
series of each brain region was extracted for subsequent
analysis. The Pearson correlation coefficient was used to
estimate the internode functional connections between the
nodal time courses. Similar to many previous dynamic
studies, we set the window length to 50 seconds and the
sliding step to one repetition time.15,17 We used a multi-
layer community detection algorithm to divide the brain
network into communities,18 and then a multilayer net-
work model was used to describe the time-varying charac-
teristics of the whole brain layer by layer.17 Each node in
one network layer is connected to itself in the other layers,
thus providing a “temporal link” or connectivity between
adjacent time points. The multilayer network model can
optimize the module division of nodes in time and space,
incorporate connectivity information from adjacent win-
dows, and depict the time flow between adjacent network
layers.17 We then calculated the recruitment and integra-
tion attributes of each brain node to quantify changes in
community structure over time based on probability and
reflect the dynamic interaction within and between differ-
ent functional systems.

1422 Volume 58, No. 5


Anatomical Network
Diffusion data preprocessing and anatomical network con-
struction were performed in the PANDA images (http://
www.nitrc.org/projects/panda/). Deterministic fiber tracking
was implemented with Diffusion Toolkit (http://trackvis.org/
dtk/) using the fiber assignment by continuous tracking
(FACT) algorithm. Terminated fiber tracking was determined
if two consecutive moving directions had a crossing angle
above 45 , or if the fractional anisotropy (FA) was out of the
threshold 0.2–1. All brain regions (90 according to AAL)
were used as nodes for tractography. The edges of the net-
work were defined as the FA values of the connected fibers
between two nodes.19 Consequently, a 90
90 adjacency
matrix representing the anatomical network for each subject
was obtained.
Morphological Network
Brain tissue segmentation and morphological network con-
struction20 were performed using the CAT12 (http://www.
neuro.uni-jena.de/cat/), as implemented in the MATLAB
(R2018b) Statistical Parametric Mapping analysis package
(SPM12; http://www.fil.ion.ucl.ac.uk/spm/soft-ware/spm12/
). The cerebral cortex was divided into 90 brain regions using
an AAL template.16 The subcortical volume in the data was
corrected for sex, age, and intracranial volume of the brain.
The resulting residual was then z-score transformed using
mean and standard deviation (SD) values of each region of
interest calculated from HCs (to derive the degrees of brain
morphological variations per region of interest relative to the
“across subjects” values). Finally, morphological networks
were generated using structural partition (n = 90) as the
nodes, and the joint variation measures between 90 morpho-
logical features were defined as the weights of the edges. Con-
sequently, for each participant, the morphological network
was represented by a symmetric weighted 90
90 matrix.
Feature Selection and MDD Diagnosis
First, the extracted features are standardized to remove the
unit limit of each feature and reduce the computational com-
plexity of the model. Then, we screened the features with
significant differences in functional, morphological, and ana-
tomical networks between patients and HCs. Age and gender
were treated as covariates in all analyses. Finally, all results
were corrected using false discovery rate (FDR) correction.
To further integrate the complementary information
from different sequences, we fused the three selected brain
network features and used the least absolute shrinkage and
selection operator (LASSO) algorithm to select the key fea-
tures most closely related to MDD from the different fea-
tures. The optimal penalization coefficient alpha was set by
10-fold cross-validation and 27 features with non-zero coeffi-
cients within the training set were selected for subsequent
analysis. Discriminant analysis was then performed using
November 2023
15222586, 2023, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jmri.28650 by Manager Information Resources Bond University Library, Wiley Online Library on [14/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Li et al.: Diagnosis of Major Depressive Disorder
random forest (RF) as a classifier. RF is a classifier that uses
multiple trees to train and predict samples. It contains multi-
ple decision trees to produce highly accurate classification
results. In the process of model classification, 10-fold cross-
validation was employed to train and improve generalization
ability. Specifically, the dataset was divided into 10 parts and
took turns to take nine of them as training data and one as
test data. This process was repeated 10 times to ensure that
each subset could be used as a test dataset. For robustness
and to reduce overfitting, we performed 10-fold cross-
validation across 100 iterations and calculated the average of
the training results over 100 iterations. Finally, we trained the
single and multisequence MRI neuroimaging features models
to compare the judgment effects of diagnosing patients with
MDD. Moreover, we input the same features to the SVM
classifier to verify the stability of previously discovered neuro-
imaging features. As mentioned earlier, the purpose of this
study was to integrate the three network features to achieve a
higher classification performance. For a better comparison,
we summarized the overall classification performance of
recent MDD studies using single and multisequence MRI.
Identify Classification Feature Weights and Predict
MDD Suicide Risk
The contribution of features in the classification model was
evaluated by calculating the weight of each feature in the
cross-validation. Finally, we identified the top 10 neuroimag-
ing features with the largest weights as baseline predictors to
explore the relationship between multisequence neuroimaging
features and suicide risk in patients with MDD. The suicidal
risk of patients with MDD was assessed according to the
Nurses Global Assessment of Suicide Risk (NGASR).21
Thirty-seven patients were classified into the high suicide risk
group (NGASR ≥ 12), whereas 35 patients were assigned to
the low suicide risk group (NGASR < 6). In the high-risk
group, all the patients had suicidal ideation and past suicide
attempts. In the low-risk group, the patients have never
attempted suicide. We then used the top 10 most discrimi-
nating neuroimaging features to predict the NGASR score
for MDD. In the model prediction process, 10-fold cross-
validation was adopted to improve the accuracy of the
model prediction, and the correlation between the
predicted and true scores was calculated to evaluate the
prediction accuracy.
Statistical Analysis
Demographic data were analyzed using Statistical Package for
Social Sciences software (version 26.0; IBM SPSS Statistics
for Windows, NY, Armonk, USA). We conducted a chi-
square (χ 2) test and a two-sample t-test to examine demo-
graphic data (age, sex, body mass index, and education level)
and Hamilton Depression Scale (HAMD) scores between the
1423

Journal of Magnetic Resonance Imaging
patients and control group. Demographic data values are
expressed as mean  standard deviation.
Two-sample t-test was performed to screen for features
with significant differences in functional networks between
patients and HCs (P < 0.05). Moreover, network-based statis-
tical analysis (NBS) was used to identify altered connectivity
in patients with MDD anatomical and morphological net-
works.22 First, we set the threshold to P < 0.01 to find these
suprathreshold connections. Next, the statistical significance
of the size of each observed component in the network was
calculated using the nonparametric permutation method
(1000 permutations). Suprathreshold connections that were
significant at a corrected level of P < 0.01 were reported.
The performance of RF and SVM classifiers was evalu-
ated using receiver operating characteristic (ROC) curves and
area under the curve (AUC), sensitivity, specificity, and accu-
racy to evaluate the diagnostic efficacy of the model. The
Delong test was used to compare the AUC of the model. In
addition, a two-sample t-test was used to compare the differ-
ence in NGASR scores between patients and HCs (P < 0.05
indicating statistical significance).
Results
Demographic Data
Details of the demographic data are presented in Table 1.
There were no significant differences between the two groups
with respect to age (t = 0.923, P = 0.498), gender
(t = 0.135, P = 0.713), and education level (t = 1.237,
P = 0.07). The MDD group had a significantly higher total
HAMD score (t = 0.607) than the HCs.
Network Analysis
Patients with MDD demonstrated remarkably distinct struc-
tural and functional aberrance compared with HCs. In the
functional network, the brain regions with two attributes
(recruitment and integration) difference in patients with
MDD were mainly located in the visual network (VIS), cog-
nitive control network (CC), auditory network (AUD), and
TABLE 1. Demographic and Clinical Characteristics (Mean  SD)
MDD (n = 139)
Age (year) 30.40  4.38
Education (year) 13.98  3.27
Sex (male/female) 49/90
BMI 11.96  3.23
HAMD score 28.76  10.37
MDD = major depressive disorder; HCs = healthy controls; HAMD = Hamilton Depression Scale; BMI = body mass index.
*Denotes significance.
1424
15222586, 2023, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jmri.28650 by Manager Information Resources Bond University Library, Wiley Online Library on [14/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
default mode network (DMN), particularly the VIS and CC
(Fig. 2). Compared to HCs, the connections of morphologi-
cal network differences in patients with MDD were mainly
located in the CC and subcortical networks (SC) (Fig. 3). In
the Anatomical network, the brain regions with FA differ-
ences between MDD patients and HCs were mainly located
in the CC and sensorimotor network (SM) (Fig. 4). We fused
the features of the three brain network differences and
selected 27 key features that were most closely related to
MDD through LASSO (see Supplemental Table S1).
Discriminative Analysis
Compared to the diagnosis of single sequence brain network
features, the classification performance of the model integrat-
ing the three brain network features was significantly
improved (see Supplemental Table S2). The RF model
showed the best performance in MDD diagnosis, with an
accuracy of 88.24%, a sensitivity of 94.44%, a specificity of
84.85%, and an AUC of 93.69% (Fig. 5b and Table 2).
From Fig. 5a, the classification result obtained with SVM is
similar to that obtained with the proposed RF model. The
classification effect of fusing three brain network feature sets
is obviously higher than that of single sequence feature sets.
All identified connections are displayed in Figs. 2–4.
The top 10 brain regions with the most significant contribu-
tions were selected for display in Table 3 and Fig. 6. The
average feature importance scores of these 10 features indi-
cated that the most discriminative features were located in the
VIS, CC, DMN, and SC. Among the 10 features that con-
tributed the most to the model, eight were from the structural
network and two from the functional network.
Predict the Suicide Risk of MDD Patients
As shown in Supplemental Table S3, NGASR scores of
MDD patients with suicide risk were significantly higher than
those of MDD patients without suicide risk. We performed a
regression analysis to predict the suicide risk of patients with
MDD based on the selection of the 10 features with the
HCs (n = 112) P value
30.92  4.46 0.498
14.44  2.40 0.07
37/75 0.713
14.60  3.32 0.607
6.79  5.71 <0.001*
Volume 58, No. 5

FIGURE 2: Comparison of dynamic network measurements between MDD and HCS. (a) Recruitment. (b) Integration. MDD = major
depressive disorder; HCs = healthy controls; L = left; R = right.
largest weight from the multisequence neuroimaging features.
As shown in Fig 7, regression analyses indicated that multi-
sequence brain image features had the highest positive corre-
lation between the predicted results and actual NGASR
scores (r = 0.691).
Discussion
In this prospective study, we integrated functional, morpho-
logical, and anatomical network features to diagnose MDD,
and the RF model achieved a good classification performance.
Compared with previous single and multisequence studies,
the RF classifier with multisequence neuroimaging features
had higher sensitivity and accuracy. Furthermore, the results
showed that our neuroimaging features had good stability in
comparison to the SVM classifier. In rs-fMRI, we used a
November 2023
15222586, 2023, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jmri.28650 by Manager Information Resources Bond University Library, Wiley Online Library on [14/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Li et al.: Diagnosis of Major Depressive Disorder
multilayer network of dynamic functional connections to
describe the time-varying characteristics among the layers of
brain regions, which helps us understand how dynamic inter-
actions between brain networks affect the clinical symptoms
of patients. Moreover, we found that multisequence neuroim-
aging features can predict the suicide risk of patients with
MDD, which provides important insights into the pathologi-
cal mechanism of MDD.
Compared with a single sequence study, the RF model
integrating three brain network feature sets greatly improved
the performance of MDD diagnosis, which indicates that dif-
ferent sequences of brain networks indeed carry complemen-
tary information related to the disease.5,23–25 This study
provides multidimensional feature information for diagnosing
MDD by building brain networks with different sequences.
A recent multisequence study based on traditional static
1425
 15222586, 2023, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jmri.28650 by Manager Information Resources Bond University Library, Wiley Online Library on [14/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Magnetic Resonance Imaging

FIGURE 3: Morphological network changes in MDD patients; (a) connections with significantly increased volume; (b) connections with
significantly reduced volume; navy blue, nodes in SM; wathet, nodes in CC; yellow, nodes in VIS; green, nodes in DMN; orange,
nodes in SC; red, nodes in AUD. SM = sensorimotor network; CC = cognitive control network; VIS = visual network; DMN = default
mode network; SC = subcortical network; AUD = auditory network; L = left; R = right.

connections achieved a good classification performance, but
its sensitivity was significantly lower than that of our study.13
The reason may be that there is time flow between brain
regions, which provides rich dynamic brain network informa-
tion for diagnosing MDD. In addition, the results showed
that the SVM model integrating three types of brain network
feature sets achieved the best classification performance,
which further indicates that our neuroimaging features have
good stability.
In the present study, the features of the brain network
integrating three different sequences provided more useful
information than a single sequence. In single-sequence diag-
nosis, compared with structural and anatomical networks, the
feature set of the functional network achieved the best classifi-
cation performance, which indicates that the time flow in the
brain region structure contains a large amount of biological
information and has potential diagnostic value. The accuracy
of the anatomical network feature set was relatively low, indi-
cating that the information provided by the anatomical net-
work is limited. This is consistent with several previous
studies that functional network feature sets are more impor-
tant for the diagnosis of mental diseases.26,27 In addition,
when we fused multisequence neuroimaging features to diag-
nose MDD, we found that functional and structural networks
yielded more discriminative features, further verifying
our view.
In the functional network, compared with HCs, we
found that the integration and recruitment degree of the VIS
and CC in MDD patients was significantly reduced, indicat-
ing that the network has been reconstructed, which not only
reduces connections within the system but also reduces func-
tional interactions between brain networks. Studies have

1426 Volume 58, No. 5

 15222586, 2023, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jmri.28650 by Manager Information Resources Bond University Library, Wiley Online Library on [14/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Li et al.: Diagnosis of Major Depressive Disorder

FIGURE 4: Increased brain regions of anatomical network in MDD patients; navy blue, nodes in SM; wathet, nodes in CC; yellow,
nodes in VIS; green, nodes in DMN; orange, nodes in SC; red, nodes in AUD. SM = sensorimotor network; CC = cognitive control
network; VIS = visual network; DMN = default mode network; SC = subcortical network; AUD = auditory network; L = left;
R = right.

FIGURE 5: The receiver operator characteristic (ROC) curves of different models: fMRI, DTI, sMRI and multimodal. (a) Support vector
machine (SVM) classifier, (b) random forest (RF) classifier. AUC = area under the receiver operating characteristic curve;
fMRI = functional MRI; sMRI = structural MRI; DTI = diffusion tensor imaging.

shown that an increase in flexibility scores leads to a decrease
in the recruitment value.17,28 In this regard, we believe that
the reduction in the integration score will make the network
more flexible, causing fragmentation of whole-brain networks.
This excessive flexibility indicates that the patient network
module lacks collaboration and stability, which reduces the
recruitment value of the network.29 It is worth noting that
the recruitment and integration values of the DMN and
AUD in the MDD group were significantly higher than those
of the control group, which indicates that the information
interaction ability between networks has an increasing trend.
In previous studies on network disorders in patients with
MDD, abnormalities in the DMN and AUD may affect the
emotional and cognitive activities of the brain.30,31 Therefore,
we infer that an increase in cross-network communication
ability may be a compensatory mechanism for maintaining
basic cognitive processes.
In the present study, abnormal morphological connec-
tions were mainly located in the CC and DMN. According
to recent studies, structural abnormalities in the CC have
consistently been found in MDD, especially in patients with
cognitive impairment.32,33 Menon proposed a classic triple

November 2023 1427

 15222586, 2023, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jmri.28650 by Manager Information Resources Bond University Library, Wiley Online Library on [14/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Magnetic Resonance Imaging

TABLE 2. Performance Comparison of Different Methods in MDD Classification

Method Modality ACC (%) SPC (%) SEN (%) AUC (%)
Bhaumik et al SVM fMRI 76.1 68.9 81.5 NA
Ramasubbu et al SVM fMRI 66.0 72.0 59.0 NA
Qiu et al SVM sMRI 72.0 69.0 75.0 NA
Su et al RF fMRI 75.9 82.6 42.7 62.9
Yang et al SVM sMRI + fMRI 84.0 81.0 88.0 91.0
Benedatta et al SVM sMRI + DTI 72.9 73.3 74.3 79.0
Maglanoc et al SVM fMRI + sMRI + DTI 61.9 42.9 69.9 61.9
Our proposed RF Multimodal 88.2 94.4 84.8 93.7

ACC = accuracy; SPC = specificity; SEN = sensitivity; AUC = area under the receiver operating characteristic curve; RF = random
forest; SVM = support vector machine; fMRI = functional MRI; sMRI = structural MRI; DTI = diffusion tensor imaging; NA = not
available.

TABLE 3. Top 10 Neuroimaging Features That Contribute Most to the RF Model

Features Types Region A Region B jweight valuej

fMRI Supp_Motor_Area_R NA 0.110
Hippocampus_L NA 0.053
sMRI Occipital_Sup_R Insula_R 0.061
Hippocampus_L Frontal_Sup_Medial_L 0.058
Putamen_R Frontal_Inf_Oper_R 0.051
Cuneus_L Frontal_Inf_Oper_R 0.039
Caudate_L Hippocampus_R 0.036
ParaHippocampal_R Rolandic_Oper_R 0.030
Cingulum_Post_L Hippocampus_R 0.023
Angular_L Occipital_Mid_L 0.022

jWj is the absolute value of the regression coefficient of the connection strength of regions A and B through LASSO.
LASSO = least absolute shrinkage and selection operator; fMRI = functional MRI; sMRI = structural MRI; NA = not available;
L = left hemisphere; R = right hemisphere.

network model, which emphasized that the abnormal connec-
tion between the CC and DMN played an important role in
switching networks and further led to the separation of the
networks.34 Based on this model, we infer that the increased
flexibility of the CC in MDD patients may lead to changes
in DMN morphological connections, which in turn leads to
cognitive impairment and low mood. Furthermore, our study
shows that the altered anatomical connections in patients
with MDD mainly involve the CC and the VIS. Compared
with HCs, the anatomical connection of patients with MDD
was significantly enhanced, indicating that the integrity of
white matter fibers in patients with MDD may be damaged
in the early stage of the disease. The CC and VIS are the
main fiber bundles that connect the contralateral homologous
brain system.35,36 We inferred that local information
processing between the hemispheres of patients with MDD
may be interrupted. However, whether this conjecture is cor-
rect requires further investigation.
In addition, the results showed that multisequence neu-
roimaging features can accurately predict suicide risk in
patients with MDD. In our study, the predicted scores were
higher than the scale’s scores. We speculate that the suicide
risk of patients may be higher than previously thought and
recommend lowering the suicide risk score threshold.

1428 Volume 58, No. 5


FIGURE 6: Subnetworks that contribute the largest 10 neuroimage features in the random forest (RF) classifier. Red, CC; green, SC;
orange, VIS; and blue, DMN. CC = cognitive control network; SC = subcortical network; VIS = visual network; DMN = default mode
network; L = left; R = right.
Of course, longitudinal studies of patients are required to
further evaluate the severity of suicide risk in individuals
with MDD. In the RF model, we found that the striatum
(putamen, caudate nucleus, and globus pallidus) and the
frontal lobe were among the areas with the greatest contri-
bution. The abnormal projection of 5-hydroxytryptamine
(5-HT) to the striatum in the midbrain limbic system
hypothesis may lead to clinical symptoms such as impulsiv-
ity and self-mutilation.37,38 Previous studies have suggested
that structural and functional changes in the frontal lobe,
amygdala, and hippocampus are the neurobiological bases
of suicide vulnerability in patients with MDD.39,40 In our
study, the neuroimaging features that contributed most to
the classification performance of the RF classifier were
mainly located in the CC, VIS, DMN, and SC (frontal
lobe, striatum, hippocampus, and cingulate gyrus), which
is consistent with previous studies.38,40
FIGURE 7: Prediction of suicide risk of MDD patients based on
the top 10 neuroimaging features with the largest feature
weight in the random forest (RF) classifier. MDD = major
depressive disorder
November 2023
15222586, 2023, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jmri.28650 by Manager Information Resources Bond University Library, Wiley Online Library on [14/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Li et al.: Diagnosis of Major Depressive Disorder
Limitations
First, the topological organization of brain networks is
affected by various parcellation strategies. Different
parcellation schemes should be used in conjunction in the
future. Second, the results of our model were obtained from a
single data set. An external data set is needed to verify these
results in the future. Third, this study is a cross-sectional
study. Future research can use multisequence brain image fea-
tures of longitudinal data for research.
Conclusion
Different sequences of brain networks carry complementary
information related to MDD. In addition, multisequence
brain image features as baseline predictors can predict the sui-
cide risk of patients with MDD, which is of great importance
for understanding the neuropathological mechanism
of MDD.
Acknowledgments
The authors thank everyone who contributed to this research.
Author Contribution
Qinghe Li, Fanghui Dong: Investigation, Data curation, For-
mal analysis, Writing - original draft. Qun Gai, Kaili Che:
Investigation, Conceptualization, Data curation, Formal anal-
ysis, Writing - original draft. Heng Ma, Feng Zhao: Data
curation. Tongpeng Chu, Ning Mao, Peiyuan Wang:
Methodology, Investigation, Writing - review and editing.
1429

Journal of Magnetic Resonance Imaging
References
1. Zhang J, Wang J, Wu Q, et al. Disrupted brain connectivity networks in
drug-naive, first-episode major depressive disorder. Biol Psychiatry
2011;70(4):334-342.
2. Ren X, Yu S, Dong W, Yin P, Xu X, Zhou M. Burden of depression in
China, 1990-2017: Findings from the global burden of disease study
2017. J Affect Disord 2020;268:95-101.
3. Holma KM, Haukka J, Suominen K, et al. Differences in incidence of
suicide attempts between bipolar I and II disorders and major depres-
sive disorder. Bipolar Disord 2014;16(6):652-661.
4. Kim JH, Suh SI, Lee HJ, Lee JH, Lee MS. Cortical and subcortical gray
matter alterations in first-episode drug-naive adolescents with major
depressive disorder. Neuroreport 2019;30(17):1172-1178.
5. Qiu L, Huang X, Zhang J, et al. Characterization of major depressive
disorder using a multiparametric classification approach based on high
resolution structural images. J Psychiatry Neurosci 2014;39(2):78-86.
6. Davis AD, Hassel S, Arnott SR, et al. White matter indices of medication
response in major depression: A diffusion tensor imaging study. Biol
Psychiatry Cogn Neurosci Neuroimaging 2019;4(10):913-924.
7. Peng X, Lau WKW, Wang C, Ning L, Zhang R. Impaired left amygdala
resting state functional connectivity in subthreshold depression individ-
uals. Sci Rep 2020;10(1):17207.
8. Koutsouleris N, Kambeitz-Ilankovic L, Ruhrmann S, et al. Prediction
models of functional outcomes for individuals in the clinical high-risk
state for psychosis or with recent-onset depression: A multimodal multi-
site machine learning analysis. JAMA Psychiatry 2018;75(11):1156-
1172.
9. Yang J, Yin Y, Zhang Z, et al. Predictive brain networks for major
depression in a semi-multimodal fusion hierarchical feature reduction
framework. Neurosci Lett 2018;665:163-169.
10. Sun K, Liu Z, Chen G, et al. A two-center radiomic analysis for differen-
tiating major depressive disorder using multi-modality MRI data under
different parcellation methods. J Affect Disord 2022;300:1-9.
11. Han K-M, De Berardis D, Fornaro M, Kim Y-K. Differentiating between
bipolar and unipolar depression in functional and structural MRI stud-
ies. Prog. Neuropsychopharmacol 2019;91:20-27.
12. Hutchison RM, Womelsdorf T, Allen EA, et al. Dynamic functional con-
nectivity: Promise, issues, and interpretations. Neuroimage 2013;80:
360-378.
13. Maglanoc LA, Kaufmann T, Jonassen R, et al. Multimodal fusion of
structural and functional brain imaging in depression using linked inde-
pendent component analysis. Hum Brain Mapp 2020;41(1):241-255.
14. Chao-Gan Y, Yu-Feng Z. DPARSF: A MATLAB toolbox for “pipeline”
data analysis of resting-state fMRI. Front Syst Neurosci 2010;4:13.
15. Allen EA, Damaraju E, Plis SM, Erhardt EB, Eichele T, Calhoun VD.
Tracking whole-brain connectivity dynamics in the resting state. Cereb
Cortex 2014;24(3):663-676.
16. Tzourio-Mazoyer N, Landeau B, Papathanassiou D, et al. Automated
anatomical labeling of activations in SPM using a macroscopic anatomi-
cal parcellation of the MNI MRI single-subject brain. Neuroimage 2002;
15(1):273-289.
17. Cui X, Ding C, Wei J, et al. Analysis of dynamic network
reconfiguration in adults with attention-deficit/hyperactivity disorder
based multilayer network. Cereb Cortex 2021;31(11):4945-4957.
18. Mucha PJ, Richardson T, Macon K, Porter MA, Onnela JP. Community
structure in time-dependent, multiscale, and multiplex networks.
Science 2010;328(5980):876-878.
19. Shu N, Liu Y, Li K, et al. Diffusion tensor tractography reveals disrupted
topological efficiency in white matter structural networks in multiple
sclerosis. Cereb Cortex 2011;21(11):2565-2577.
20. Yun JY, Boedhoe PSW, Vriend C, et al. Brain structural covariance net-
works in obsessive-compulsive disorder: A graph analysis from the
ENIGMA consortium. Brain 2020;143(2):684-700.
1430
15222586, 2023, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jmri.28650 by Manager Information Resources Bond University Library, Wiley Online Library on [14/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
21. Cutcliffe JR, Barker P. The Nurses’ global assessment of suicide risk
(NGASR): Developing a tool for clinical practice. J Psychiatr Ment
Health Nurs 2004;11(4):393-400.
22. Tyan YS, Liao JR, Shen CY, Lin YC, Weng JC. Gender differences in the
structural connectome of the teenage brain revealed by generalized
q-sampling MRI. Neuroimage Clin 2017;15:376-382.
23. Su D, Zhang X, He K, Chen Y. Use of machine learning approach to
predict depression in the elderly in China: A longitudinal study.
J Affect Disord 2021;282:289-298.
24. Bhaumik R, Jenkins LM, Gowins JR, et al. Multivariate pattern analysis
strategies in detection of remitted major depressive disorder using
resting state functional connectivity. Neuroimage Clin 2017;16:
390-398.
25. Ramasubbu R, Brown MR, Cortese F, et al. Accuracy of automated clas-
sification of major depressive disorder as a function of symptom sever-
ity. Neuroimage Clin 2016;12:320-331.
26. Achalia R, Sinha A, Jacob A, et al. A proof of concept machine learning
analysis using multimodal neuroimaging and neurocognitive measures
as predictive biomarker in bipolar disorder. Asian J Psychiatr 2020;50:
101984.
27. Yang J, Zhang M, Ahn H, et al. Development and evaluation of a multi-
modal marker of major depressive disorder. Hum Brain Mapp 2018;
39(11):4420-4439.
28. Mattar MG, Cole MW, Thompson-Schill SL, Bassett DS. A functional
cartography of cognitive systems. PLoS Comput Biol 2015;11(12):
e1004533.
29. Braun U, Schafer A, Bassett DS, et al. Dynamic brain network
reconfiguration as a potential schizophrenia genetic risk mechanism
modulated by NMDA receptor function. Proc Natl Acad Sci U S A
2016;113(44):12568-12573.
30. Tian S, Zhang S, Mo Z, et al. Antidepressants normalize brain flexibility
associated with multi-dimensional symptoms in major depressive
patients. Prog Neuropsychopharmacol Biol Psychiatry 2020;100:
109866.
31. Hou Z, Sui Y, Song X, Yuan Y. Disrupted interhemispheric synchrony in
default mode network underlying the impairment of cognitive flexibility
in late-onset depression. Front Aging Neurosci 2016;8:230.
32. Pan F, Xu Y, Zhou W, et al. Disrupted intrinsic functional connectivity of
the cognitive control network underlies disease severity and executive
dysfunction in first-episode, treatment-naive adolescent depression.
J Affect Disord 2020;264:455-463.
33. Wang YL, Yang SZ, Sun WL, Shi YZ, Duan HF. Altered functional inter-
action hub between affective network and cognitive control network in
patients with major depressive disorder. Behav Brain Res 2016;298:
301-309.
34. Menon V. Large-scale brain networks and psychopathology: A unifying
triple network model. Trends Cogn Sci 2011;15(10):483-506.
35. He X, Pueraro E, Kim Y, et al. Association of white matter integrity with
executive function and antidepressant treatment outcome in patients
with late-life depression. Am J Geriatr Psychiatry 2021;29(12):1188-
1198.
36. Coloigner J, Batail J-M, Commowick O, et al. White matter abnormali-
ties in depression: A categorical and phenotypic diffusion MRI study.
NeuroImage 2019;22:101710.
37. Morgane PJ, Galler JR, Mokler DJ. A review of systems and networks
of the limbic forebrain/limbic midbrain. Prog Neurobiol 2005;75(2):
143-160.
38. Bennett MR. The prefrontal-limbic network in depression: Modulation
by hypothalamus, basal ganglia and midbrain. Prog Neurobiol 2011;
93(4):468-487.
39. Bani-Fatemi A, Tasmim S, Graff-Guerrero A, et al. Structural and func-
tional alterations of the suicidal brain: An updated review of neuroim-
aging studies. Psychiatry Res Neuroimaging 2018;278:77-91.
40. van Heeringen K, Mann JJ. The neurobiology of suicide. Lancet Psychi-
atry 2014;1(1):63-72.
Volume 58, No. 5