Content uploaded by Maziar Sabouri
Author content
All content in this area was uploaded by Maziar Sabouri on Nov 15, 2022
Content may be subject to copyright.
Vol.:(0123456789)
1 3
Journal of Digital Imaging
https://doi.org/10.1007/s10278-022-00705-9
ORIGINAL PAPER
Myocardial Perfusion SPECT Imaging Radiomic Features andMachine
Learning Algorithms forCardiac Contractile Pattern Recognition
MaziarSabouri1,2· GhasemHajianfar2· ZahraHosseini2· MehdiAmini3· MobinMohebi4· TaherehGhaedian5·
ShabnamMadadi2· FereydoonRastgou2· MehrdadOveisi6,12· AhmadBitarafanRajabi7,8· IsaacShiri3 ·
HabibZaidi3,9,10,11
Received: 21 July 2022 / Revised: 31 August 2022 / Accepted: 15 September 2022
© The Author(s) 2022
Abstract
A U-shaped contraction pattern was shown to be associated with a betterCardiac resynchronization therapy (CRT) response.
The main goal of this study is to automatically recognize left ventricular contractile patterns using machine learning algo-
rithms trained on conventional quantitative features (ConQuaFea) and radiomic features extracted fromGated single-photon
emission computed tomography myocardial perfusion imaging (GSPECT MPI). Among 98 patients with standard resting
GSPECT MPI included in this study, 29 received CRT therapy and 69 did not (also had CRT inclusion criteria but did not
receive treatment yet at the time of data collection, or refused treatment). A total of 69 non-CRT patients were employed
for training, and the29 were employed for testing. The models were built utilizing features from three distinct feature sets
(ConQuaFea, radiomics, and ConQuaFea + radiomics (combined)), which were chosen usingRecursive feature elimination
(RFE) feature selection (FS), and then trained using seven different machine learning (ML) classifiers. In addition, CRT
outcome prediction was assessed by different treatment inclusion criteria as the study’s final phase. The MLP classifier
had the highest performance among ConQuaFea models (AUC, SEN, SPE = 0.80, 0.85, 0.76). RF achieved the best perfor-
mance in terms of AUC, SEN, and SPE with values of 0.65, 0.62, and 0.68, respectively, among radiomic models. GB and
RF approaches achieved the best AUC, SEN, and SPE values of 0.78, 0.92, and 0.63 and 0.74, 0.93, and 0.56, respectively,
among the combined models. A promising outcome was obtained when using radiomic and ConQuaFea from GSPECT MPI
to detect left ventricular contractile patterns by machine learning.
Keywords Machine learning· CRT · GSPECT MPI· Radiomics· Quantitative features
* Ahmad Bitarafan Rajabi
bitarafan@hotmail.com
* Isaac Shiri
isaac.shirilord@unige.ch
* Habib Zaidi
habib.zaidi@hcuge.ch
1 Department ofMedical Physics, School ofMedicine, Iran
University ofMedical Science, Tehran, Iran
2 Rajaie Cardiovascular Medical andResearch Center, Iran
University ofMedical Science, Tehran, Iran
3 Division ofNuclear Medicine andMolecular Imaging,
Geneva University Hospital, CH-1211Geneva4, Switzerland
4 Department ofBiomedical Engineering, Tarbiat Modares
University, Tehran, Iran
5 Nuclear Medicine andMolecular Imaging Research Center,
School ofMedicine, Namazi Teaching Hospital, Shiraz
University ofMedical Sciences, Shiraz, Iran
6 Comprehensive Cancer Centre, School ofCancer &
Pharmaceutical Sciences, Faculty ofLife Sciences &
Medicine, King’s College London, London, UK
7 Echocardiography Research Center, Rajaie Cardiovascular
Medical andResearch Center, Iran University ofMedical
Sciences, Tehran, Iran
8 Cardiovascular Interventional Research Center, Rajaie
Cardiovascular Medical andResearch Center, Iran University
ofMedical Sciences, Tehran, Iran
9 Geneva University Neurocenter, Geneva University, Geneva,
Switzerland
10 Department ofNuclear Medicine andMolecular Imaging,
University ofGroningen, University Medical Center
Groningen, Groningen, Netherlands
11 Department ofNuclear Medicine, University ofSouthern
Denmark, Odense, Denmark
12 Department ofComputer Science, University ofBritish
Columbia, VancouverBC, Canada
Journal of Digital Imaging
1 3
Abbreviations
ConQuaFea Conventional quantitative features
HF Heart failure
CRT Cardiac resynchronization therapy
NYHA New York Heart Association
LVEF Left ventricular ejection fraction
LV Left ventricle
CAD Coronary artery disease
GSPECT MPI Gated single-photon emission com-
puted tomography myocardial perfusion
imaging
NCM Non-contact mapping
CMR Cardiac magnetic resonance
MRI Magnetic resonance imaging
CT Computed tomography
PET Positron emission tomography
QGS Quantitative gated SPECT
IBSI Image Biomarker Standardization
Initiative
VOI Volume of interest
SD Standard deviation
RFE Recursive feature elimination
LR Logistic regression
DT Decision tree
RF Random forest
XGB Extreme gradient boosting
MLP Multi-layer perceptron
SVM Support vector machine
GB Gradient boosting
AUC Area under the ROC curve
ACC Accuracy
SEN Sensitivity
SPE Specificity
FDR False discoveries rate
Introduction
Heart failure (HF) is a relatively common cardiovascular
disorder with prominent morbidity and mortality [1]. HF is
closely related to left ventricular (LV) cardiac mechanical
dyssynchrony, which reflects timing differences across vari-
ous left ventricle regions [2]. Nonhomogeneous contraction
patterns can be caused by the uncoordinated distribution of
electrical activation in the heart pathways, known as car-
diac dyssynchrony [3–5]. Thus, therapeutic measures are
developed to resynchronize the left ventricular contraction
in HF patients.
Cardiac resynchronization therapy (CRT) demonstrated
significant success in treating patients with fatal HF [6, 7].
Patients undergoing CRT have substantiated the claims of
most previous studies regarding the enhancements in sev-
eral parameters, such as 6-min walking distance, New York
Heart Association (NYHA) functional class, quality of life
score, and peak O2 [6]. Nonetheless, estimations demon-
strated that approximately one-third of chosen cases do not
respond to this costly and invasive therapy [8–12] in spite
of meeting current inclusion criteria for CRT therapy by the
guidelines, which includes NYHA III or IV, left ventricular
ejection fraction (LVEF) < 35%, and QRS duration ≥ 130ms
[13]. Consequently, the search for more specific criteria is
still under investigation.
In line with finding new and more appropriate indicators
for CRT patient selection, several studies have been con-
ducted. In a study by Bax etal. [14], LV dyssynchrony was
introduced as a desirable indicator. Furthermore, in another
study conducted by Chen etal. [15], histogram bandwidth
and phase standard deviation parameters extracted from
phase analysis were suggested as important indicators.
In addition, in examining the locations of CRT leads by
Adelstein etal. [9], it was found that the best locations
are far away from the scar tissue [9]. New LV mechanical
dyssynchrony parameters were extracted from GSPECT
MPI phase analysis with deep learning to aid CRT patient
selection by He etal. [16].
Two types of left ventricular contraction patterns includ-
ing U-shaped and non-U-shaped patterns have been recently
proposed by different imaging modalities [17]. A U-shaped
pattern is formed by a left ventricle–directed linear blockage
which impedes contraction propagation [18, 19]. Conversely,
a non-U-shaped pattern consists of two types, namely,
“homogenous contraction” with apparent delay on all walls
and “heterogeneous contraction” with multiple contraction
delays in different sites [19]. An improved CRT response is
discerned to be in association with a U-shaped contraction
pattern [17–22].
Myocardial perfusion imaging (MPI) with gated single-
photon emission computed tomography (GSPECT) is a
practical technique to assess perfusion and function of left
ventricle (LV). Evaluation of LV dyssynchrony and LV con-
traction patterns can also be ascertained by applying phase
analysis to GSPECT MPI. The advantage of simultaneous
assessment of perfusion abnormalities, such as the extent
and severity of ischemia and scar, and functional param-
eters including LV dyssynchrony, LVEF, and LV volumes
make this modality an eligible method for evaluation and
diagnosis of patients with HF and LV dysfunction. Besides,
the automated nature of phase analysis results in an accept-
able repeatability and reproducibility of this technique. As
a result, MPI GSPECT is a relevant modality for examining
the left ventricular contractile pattern [19, 23].
Recent studies concerning quantitative radiomics anal-
ysis, through acting as biomarkers, have provided new
insights into better handling of diseases, such as cancer [24]
and coronary artery disease (CAD) to predict survival [25,
26], prognosis [27, 28], and therapeutic response [29, 30],
Journal of Digital Imaging
1 3
different pathology classification [28, 31–33], and accumu-
lating data for personalized medicine [34]. In fact, radiomics
is an almost new science that has attracted many researchers’
attention and is therefore growing rapidly. However, it is not
yet fully prepared to enter the clinical phase. In fact, radi-
omics analysis has been performed on magnetic resonance
imaging (MRI), computed tomography (CT), and positron
emission tomography (PET) images, but less on SPECT
owing to its low spatial resolution and low sensitivity. It
should be noted that in recent studies, radiomics has per-
formed well in brain SPECT images as well as heart SPECT
images [35–37]. Radiomic features are used to feed machine
learning algorithms. Machine learning techniques, through
the usage of computer algorithms and advanced statistical
techniques, facilitate automatic extraction of prognostic
knowledge or discriminatory patterns from data, often with
the aim of making prediction on new data [38–40].
In this work, the purpose of automatic detection of left
ventricular contractile patterns is fulfilled by radiomic
and conventional quantitative features (ConQuaFea) using
machine learning algorithms and MPI GSPECT images.
This study aims to help clinicians to more confidently select
patients who are eligible for CRT treatment and will also
save them time and energy. In the last step, a comparison
was made with the evaluation of response to CRT treatment
using the left ventricular contraction pattern diagnosed by
two experienced nuclear medicine physicians and the mod-
els presented in this study and standard criteria for prescrib-
ing CRT.
Materials andMethods
Figure1 represents the graphical pipeline of the study. The
whole workflow of the study, from data acquisition to evalu-
ation of the proposed models, is elaborated in the following
sub-sections as shown in Fig.1.
Dataset andImage Acquisition
This is a retrospective study including 98 patients encom-
passing 29 patients who underwent CRT treatment and
69 patients who did not (not being treated yet at the time
of data collection or refused treatment) but had the same
inclusion criteria: (a) EF ≤ 35% based on echocardiog-
raphy and (b) QRS duration ≥ 130ms [13]. Eighty-three
men and 15 women (mean age = 59.37) were selected as
participants. Out of 98 patients, 48 had U-shaped and 50
had non-U-shaped contractile patterns (visually assessed
by two experienced nuclear medicine physicians from the
polar maps as described in detail in the following sec-
tion). Overall, among the patients who underwent CRT,
13 had U-shaped and 16 had non-U-shaped patterns, and
among those patients who did not undergo CRT, 35 had
U-shaped and 34 had non-U-shaped patterns. All patients
underwent MPI GSPECT with the same vendor and same
acquisition parameters. Each patient registered, underwent
conventional resting MPI GSPECT. At rest condition,
555–740MBq of Tc-99m sestamibi was intravenously
administered, and the GSPECT scan started 45–90min
Fig. 1 Flowchart describing the main steps involved in the presented study
Journal of Digital Imaging
1 3
post-injection. The images were acquired on a dual-headed
gamma camera (Symbia™ T2, Siemens Healthcare) with
body auto-contour form 135° (RAO) to − 45° (LAO) in
180° orbit, 32 thirty-second steps and 16-bin gating,
using a matrix size of 64 × 64 with an isotropic voxel size
(6.591 × 6.591 × 6.591 mm3) in the reconstructed images.
The photopeak was adjusted to 140keV with 20% energy
window. Image reconstruction was accomplished using fil-
tered backprojection technique with a post-reconstruction
Butterworth filter (order = 5, cut-off frequency of 0.45
cycles/mm).
Conventional Quantitative Features (ConQuaFea)
Phase analysis was also performed using quantitative gated
SPECT (QGS) software (Table1) to extract several quantita-
tive image-based features. These are summarized in Table1,
“Features extracted from Quantitative Gated SPECT (QGS)”
Table 1 First part of ConQuaFea including phase analysis and QGS features, separately for cohorts of patients with U-shaped and non-U-shaped
contractile patterns. IQR interquartile range
ConQuaFea first part Non-u-shaped (n = 50)
Mean and IQR U-shaped (n = 48)
Mean and IQR
P value
Age (years) 59.52 and 13.00 59.23 and 13.25 0.070
Ejection fraction (Echo)
(%) 29.16 and 10.00 30.10 and 10.25 0.269
Phase analysis indices Apex Bandwidth (ms) 125.20 and 114.00 128.63 and 114.00 0.754
Mean 356.32 and 88.00 340.50 and 86.25 0.478
Standard deviation 35.70 and 35.00 37.96 and 35.00 0.947
Entropy (%) 39.11 and 33.90 39.51 and 34.03 0.535
Lateral Bandwidth (ms) 123.70 and 104.00 164.73 and 112.50 0.018
Mean 340.46 and 78.00 368.58 and 78.50 0.755
Standard deviation 35.30 and 27.00 46.96 and 30.25 0.032
Entropy (%) 45.87 and 18.80 48.06 and 19.43 0.134
Inferior Bandwidth (ms) 121.88 and 93.00 143.92 and 92.25 0.523
Mean 341.10 and 59.00 363.38 and 56.75 0.047
Standard deviation 34.24 and 27.00 41.46 and 27.50 0.395
Entropy (%) 46.97 and 17.40 49.02 and 17.15 0.723
Septal Bandwidth (ms) 134.96 and 118.00 147.71 and 116.00 0.476
Mean 344.86 and 78.00 350.96 and 78.75 0.138
Standard deviation 38.84 and 35.00 40.33 and 34.25 0.981
Entropy (%) 47.40 and 19.60 45.60 and 19.70 0.941
Anterior Bandwidth (ms) 131.54 and 116.00 159.21 and 114.75 0.301
Mean 333.36 and 57.00 350.90 and 57.00 0.288
Standard deviation 36.40 and 35.00 43.60 and 34.50 0.312
Entropy (%) 46.59 and 20.80 50.07 and 20.35 0.864
Features extracted from
quantitative gated SPECT
(QGS)
Lung heart ratio 0.36 and 0.11 0.40 and 0.11 0.155
Summed motion score 27.00 and 20.00 25.90 and 19.25 0.000
Summed thickening score 18.26 and 16.00 18.10 and 14.50 0.004
Summed motion (%) 31.74 and 23.00 30.46 and 22.25 0.000
Summed thickening (%) 35.72 and 31.00 35.46 and 29.50 0.004
End-diastolic volume (QGS) (ml) 180.20 and 80.00 153.27 and 80.75 0.076
End systolic volume (QGS) (ml) 123.68 and 68.00 98.94 and 67.50 0.032
Systolic volume (ml) 56.80 and 26.00 54.29 and 26.00 0.732
Ejectionfraction (%) 37.02 and 17.00 37.81 and 17.00 0.030
Peak emptying rate (EDV/s) − 2.07 and 1.04 − 1.99 and 0.99 0.181
Peak filling rate (EDV/s) 1.44 and 0.85 1.72 and 0.85 0.594
Peak filling rate2 (EDV/s) 1.39 and 0.97 1.00 and 0.98 0.243
Mean filling rate/3 (EDV/s) 0.74 and 0.60 0.84 and 0.55 0.866
Time to peak filling from ES (ms) 138.25 and 80.00 159.56 and 79.50 0.914
Beats per minute (beats/minute) 773.20 and 224.00 856.54 and 235.50 0.788
Journal of Digital Imaging
1 3
section. Furthermore, other categorical image-based fea-
tures, such as perfusion, phase, and wall motion, were also
collected from polar maps (Table2). In addition, explana-
tions related to the phase analysis and QGS features are pro-
vided in Supplementary Table A.1. The results report the
mean and interquartile range (IQR). The p values were cal-
culated for continuous features using t-test and for categori-
cal features using chi-square. Moreover, the way of scoring
different regions of the heart is available in Table3 accord-
ing to [41]. Quantitative image-based and phase analysis
data were combined and are reported as conventional quanti-
tative features (ConQuaFea) set in the rest of the manuscript.
Investigation ofMyocardial Contractile Patterns
fromPolar Maps
The myocardial uptake is divided into a limited number
of raw perfusion samples after the left ventricle has been
segmented, each represented by the mean or maximal wall
photon counts at a specific place on the myocardial surface
[42]. The objective is to parametrically represent myocar-
dial perfusion to enable standard inter-subject comparability.
A visual representation known as a “polar map” or “bull’s
eye map” was created to make it easier to analyze the data
obtained by polar sampling [43]. In clinical practice, polar
maps are often employed because they provide a fast vis-
ual overview of myocardial perfusion data for the whole
LV. Additionally, polar maps enable the myocardium to be
divided into sections defined by vascular regions, LV wall,
normalized 21-segment, 17-segment, or 5-segment status,
which are useful for localizing the abnormalities [44, 45].
Mechanical left ventricular dyssynchrony can produce
different patterns in polar map images as a result of two
types of electrical activity propagation in the left ventricu-
lar myocardium, which include a U-shaped pattern and a
non-U-shaped pattern. U-shaped pattern means that elec-
trical activity propagation is blocked in a line while the
other walls are almost contracted. For example, a contrac-
tion of an area starts from the anteroseptal wall and then
goes to the apex, the lateral wall, and finally to the anterior
wall. However, there are two types of non-U-shaped pat-
tern. The first type is a homogeneous pattern; for example,
contraction goes from the septal wall to the lateral wall.
The second type is the heterogeneous pattern, in which
there are multiple areas for myocardial contraction [19].
Different contraction patterns of the left ventricle can be
seen in Fig.2. In this study, the contraction pattern was
assessed visually by two nuclear medicine physicians from
the 17-segment mode polar maps polar maps.
Table 2 Second part of
ConQuaFea including wall
motion, phase, and perfusion
features, separately for cohorts
of patients with U-shaped
and non-U-shaped contractile
patterns
ConQuaFea second part Non-u-shaped U-shaped P value
Wall motion Septal # in classes (0, 1, 2, 3, 4, 5) 2, 3, 8, 7, 10, 20 7, 4, 9, 5, 14, 9 0.150
Anterior 16, 7, 11, 5, 10, 1 17, 6, 11, 10, 3, 1 0.357
Lateral 19, 8, 9, 10, 3, 1 22, 7, 10, 6, 3, 0 0.806
Inferior 10, 8, 8, 9, 13, 2 14, 4, 9, 7, 13, 1 0.761
Apex 11, 6, 6, 7, 11, 9 10, 3, 7, 2, 18, 8 0.346
Phase Septal # in classes (1, 2, 3, 4, 5) 9, 11, 8, 9, 13 10, 7, 11, 8, 12 0.831
Anterior 12, 13, 10, 12, 3 11, 14, 11, 9, 3 0.972
Lateral 6, 12, 12, 8, 12 11, 10, 5, 9, 13 0.331
Inferior 9, 4, 14, 16, 7 5, 9, 8, 16, 10 0.268
Apex 17, 9, 6, 4, 14 10, 11, 14, 4, 9 0.180
Perfusion Septal # in classes (0, 1, 2, 3, 4) 28, 10, 4, 6, 2 31, 6, 4, 4, 3 0.788
Anterior 33, 3, 9, 3, 2 33, 6, 3, 3, 3 0.385
Lateral 38, 6, 2, 3, 1 38, 5, 4, 1, 0 0.606
Inferior 26, 8, 8, 6, 2 29, 11, 5, 3, 0 0.368
Apex 20, 7, 5, 8, 10 26, 4, 3, 5, 10 0.600
Table 3 Scoring heart segments in terms of wall motion, perfusion,
and phase
Categorical features Scores
Wall motion 0 = normal
1 = mild hypokinesia
2 = moderate hypokinesia
3 = severe hypokinesia
4 = akinesia
5 = dyskinesia
Perfusion 0 = normal
1 = mild hypoperfusion
2 = moderate hypoperfusion
3 = severe hypoperfusion
4 = absent perfusion
Phase 1 = fastest contraction
5 = slowest contraction
Journal of Digital Imaging
1 3
Radiomics Feature Extraction
To extract MPI SPECT radiomic features, the left ventricle
was segmented by an experienced nuclear medicine tech-
nologist and edited/verified by a nuclear medicine physician.
Then, feature extraction was conducted utilizing pyradiom-
ics, a library in python compliant with image biomarker
standardization initiative (IBSI) [40, 46]. Re-sampling was
performed for all images with order 3 interpolation using
sitkBSpline to 6.591 × 6.591 × 6.591 mm3 voxels. Intensi-
ties within the volume of interest (VOI) were separated to
32 discrete gray levels with fixed bin number technique. A
total of 107 radiomic features including shape, intensity, and
second-/high-order texture features from GLDM, GLCM,
GLRLM, GLSZM, and NGTDM families were extracted.
The name and short description of all radiomic features are
presented in Supplementary Table A.2.
Machine Learning Workflow
As can be seen in Fig.1, after the feature extraction process,
the data was split into train/test partitions in a way that
69 non-CRT patients were used for training and 29 CRT
patients were used for testing. In addition, three feature
sets were pursued for modeling. In the first case, only Con-
QuaFea was used for modeling. In the second case, only
radiomic features and finally, in the third case, a combi-
nation of these two feature sets (combined) were used for
modeling. In all models, the features extracted from the
training dataset were normalized using Z-score method,
and the calculated mean and standard deviation (SD) were
applied on corresponding features extracted from the test
dataset.
The feature selection method used in this study was recur-
sive feature elimination (RFE). Modeling was performed
using logistic regression (LR), decision tree (DT), random for-
est (RF), extreme gradient boosting (XGB), multi-layer per-
ceptron (MLP), support vector machine (SVM), and gradient
boosting (GB) algorithms. As a result, a total of 21 different
models (3 kinds of feature-sets including ConQuaFea, radiom-
ics, and combined × 1 feature selection method × 7 machine
learning methods) were implemented. Hyper-parameters were
optimized using GridSearch with 5-fold cross-validation in
training data, and best values were selected to train model
followed by applying the trained model on test data by 1000
bootstrap. The best hyper-parameters for each classifier are
presented in Table A.3. Area under the ROC curve (AUC),
accuracy (ACC), sensitivity (SEN), and specificity (SPE)
Fig. 2 Illustration of different
contraction patterns of the left
ventricle showing U-shaped
(upper right), non-U-shaped
(upper left) (homogeneous), and
non-U-shaped (heterogeneous)
(bottom)
Journal of Digital Imaging
1 3
metrics were used to evaluate the models. All analysis was
performed on Python 3.8.8.
Delong Test
Using Delong test, comparisons between AUC of all models
were performed, which was in turn followed by false dis-
coveries rate (FDR) correction with Benjamini–Hochberg
method applied on p values. Consequently, adjusted p values,
also known as q values, were assessed. P values less than
0.05 were considered statistically meaningful.
Evaluation ofCRT Response
In the final step of the study, the task of differentiating
between patients responding to CRT or not was assessed
by different inclusion criteria for the treatment, namely, (1)
conventional criteria and (2) left ventricle contractile pattern
status assessed visually by two nuclear medicine physicians
or alternatively by machine learning and ConQuaFea/radi-
omic features. These patients considered as responders to
CRT, experienced an improvement of at least 5% in LVEF
during six months of follow-up [13].
Results
Selected Features
Based on RFE feature selection, 9 ConQuaFea, 18 com-
bined, and 12 radiomic features were selected. Figure3
illustrates the distribution of selected radiomic features
over different feature families in the form of a pie chart.
As it is shown in Fig.3, GLCM and GLSZM features
were selected the most, following them were GLDM and
NGTDM which had the highest selection. This is while,
first order in radiomic features and GLRLM in com-
bined features were not selected by the feature selection
methods.
Figure4 illustrates the selection process and lists the
selected features by RFE feature selection for radiomics,
ConQuaFea, and combined models.
Models’ Performance
The models were developed using features from three dif-
ferent feature sets (ConQuaFea, radiomics, and combined),
selected by RFE feature selection method, and trained with
seven different machine learning methods (LR, DT, RF,
XGB, MLP, SVM, and GB). Figure5 illustrates perfor-
mance metrics of all models. The performance metrics
reported in Fig.5 include AUC, accuracy, sensitivity, and
specificity.
The best performance among ConQuaFea models was
achieved by MLP classifier (ACC, AUC, SEN, SPE = 0.80,
0.80, 0.85, 0.76, respectively). Among radiomics models the
best models was RF model (ACC, AUC, SEN, SPE = 0.66,
0.65, 0.62, 0.68, respectively). Among the combined mod-
els, the best ones were GB and RF machines (ACC, AUC,
SEN, SPE were 0.76, 0.78, 0.92, 0.63, and 0.72, 0.74, 0.93,
0.56, respectively).
The results of the Delong test are illustrated in Fig.6. The
AUC of each model was compared with 20 other models.
The results were classified as statistically significant (sig-
nificantly lower or significantly higher) and non-significant.
In Fig.6, the MLP classifier on ConQuaFea had the best
results with 11 significantly higher q values. Regarding radi-
omic features, the RF classifier did not have any significant
q values. In terms of models with combined feature sets,
GB and RF classifiers had 11 and 3 significantly higher q
values, respectively.
Fig. 3 Distribution of selected radiomic features by RFE feature
selection method across different feature families. The left chart is
related to the time when the features were selected from the radiomic
data set, whereas the right chart is related to the time when the radi-
omic features were selected from combined features
Journal of Digital Imaging
1 3
CRT Response Prediction
The primary aim of this study was to predict contractile
patterns using GSPECT MPI, but the performance of the
different models regarding the prediction of CRT response
for 29 patients undergoing the treatment were also evalu-
ated. Table4 illustrates outcomes of 29 patients who
underwent CRT with the confusion matrix regarding the
prediction of models based on conventional criteria and
myocardium contractile pattern (identified by two nuclear
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
4
7
10
13
16
19
22
25
28
31
34
37
40
43
46
49
52
55
58
61
64
67
70
73
76
79
82
85
88
91
94
97
100
103
106
109
112
115
118
121
124
127
130
133
136
139
142
145
148
151
154
157
160
RFE Feature Selection
ConQuaFea Radiomics Combined Selected Point
ConQuaFea APX_Mean, APX_stDev, INF_stDev, ANT_Mean, LHR, SV, TTPF, EF_Echo, Gender
Radiomics Shape_Sphericity, GLDM_GLV, GLDM_GLNU, GLCM_JEnergy, GLCM_IV,GLCM_CP, GLRLM_GLV,
GLSZM_SZNUN, GLSZM_SZNU, GLSZM_SAHGLE, NGTDM_Coarseness, NGTDM_Strength
Combined
Shape_Sphericity,GLDM_GLNU, GLCM_CS, GLCM_IMC1, FO_Skewness,GLSZM_SZNUN, GLSZM_SZNU,
NGTDM_Strength, APX_Mean, APX_Entropy_Percent, LAT_Bandwidth, ANT_Mean,LHR, STS, SM_Percent, SV,
BPM, INF_Per
ycaruccA
Variables
Fig. 4 The process of RFE feature selection and the resultant features.
The green circles show the selected points (APX_Mean, apex mean;
APX_stDev, apex standard deviation; INF_stDev, inferior stand-
ard deviation; ANT_Mean, anterior mean; LHR, lung heart ratio;
SV, systolic volume; TTPF, time to peak filling from ES; EF_Echo,
ejection fraction (Echo); APX_Entropy_Percent, apex entropy (%);
LAT_Bandwidth, lateral bandwidth; STS, summed thickening score;
SM_Percent, summed motion (%); BPM, beats per minute; INF_Per,
inferior perfusion)
Fig. 5 Performance metrics
of ConQuaFea, radiomic, and
combined features. The metrics
include the area under the curve
(AUC), accuracy (ACC), sen-
sitivity (SEN), and specificity
(SPE). Seven different classi-
fiers with a feature selection
method and 3 different feature
sets are considered
Journal of Digital Imaging
1 3
medicine physicians and different proposed machines) for
prescription of the treatment. Table4 is color coded as blue
for correct decisions (true positive and negative) and red
for wrong decisions (false positive and negative). The first
row of Table4 shows that based on the conventional crite-
ria, from 29 patients prescribed with CRT, only 16 patients
Fig. 6 Comparison of model’s performance through the Delong test
being applied on the AUCs of models. We compare the pairwise
model in this figure where the row models were tested against column
models. Light blue, if the row model had significantly higher p value
than the column model; purple, if the row model had significantly
lower p value compared to the column model; red, if the compari-
son between the row model and column model had non-significant p
value
Table 4 Prediction of models based on conventional criteria and myocardium contractile pattern (identified by two nuclear medicine physicians
and different proposed machines) for the prescription of treatment
CRT selection Criteria
Contractile Pattern
Status Responders
Non-
responders
Success
Rate
Conventional Criteria _ 16 13 55%
Contractile Pattern identified
by human (ground truth)
U-shaped 11 2 76%
Non-U-shaped 5 11
Contractile Pattern identified
by Combined_RFE_SVM
U-shaped 11 5 66%
Non-U-shaped 5 8
Contractile Pattern identified
by Combined_RFE_MLP
U-shaped 8 4 59%
Non-U-shaped 8 9
Contractile Pattern identified
Radiomics_RFE_MLP
U-shaped 7 3 59%
Non-U-shaped 9 10
Contractile Pattern identified
by ConQuaFea_RFE_MLP
U-shaped 8 5 55%
Non-U-shaped 8 8
Journal of Digital Imaging
1 3
(55%) responded positively after the treatment whereas 13
other patients (45%) failed in treatment. The second row
shows the scenario in which the treatment criteria was set
as the contractile pattern statue (identified by humans). In
this case, 13 patients were identified with U-shaped pattern,
supposed to be sent for treatment, and if so, only 2 of them
would have failed. In addition, from the 16 patients which
were identified as non-U-shaped, they were not supposed
to be sent for treatment, 5 of them would have responded
positively to the treatment. The following rows (identified
by machine) can be seen in Table4.
Discussion
Echocardiography, non-contact mapping (NCM), cardiac
magnetic resonance (CMR), and myocardial perfusion imag-
ing (MPI) are discovering two types of contraction patterns
in the left ventricle. NCM is considered the gold standard
of assessing the patterns of LV electrical activation. Yet the
hazardous and invasive nature of the procedure restricts its
wide clinical application [18]. Accessibility, noninvasive-
ness and low cost are the advantages leading to wide usage
of echocardiography. However, low repeatability of echo-
based parameters, the state of being operator-dependent and
suboptimal acoustic window for 20% of the patients have
led to the approach being unpromising in selecting patients
for CRT [52, 53]. CMR has emerged as a valuable tool to
assess LV contraction patterns due to its high resolution and
excellent tissue characterization which make it a promising
approach for selecting responsive patients to CRT. Never-
theless, a high percentage of patients are disqualified on
account of having pacemakers and implantable cardioverter
defibrillators (ICDs) or as a result of being claustrophobic.
In addition, the procedure is expensive, difficult to access,
and extensively time-consuming [52–54]. GSPECT MPI
transpires to be a practical technique to ascertain LV con-
traction patterns. Furthermore, GSPECT MPI is consider-
ably being used for HF patients to identify LV dyssynchrony,
LVEF, LV volumes, ischemia, viability, and scar tissue.
Moreover, the advantages include the state of being ubiqui-
tous and automated [19, 23, 49].
To the best of our knowledge, our study represents the
first attempt to label left ventricular contractile patterns by
ConQuaFea and radiomic features using machine learning
and GSPECT MPI images. While this task has been previ-
ously studied in GSPECT MPI [19], it has not been investi-
gated by machine learning approaches as well as radiomic
features. In this study, we applied multiple machine learning
and a feature selection method on feature sets, including
radiomics, ConQuaFea/phase-analysis, and combination of
both features to develop machines for the identification of
LV contractile pattern from the MPI GSPECT images.
Toward the identification of most relative radiomic fea-
tures, Fig.3 illustrates the distribution of radiomic features
selected by RFE feature selection algorithm over radiomic
feature families. As it can be seen, GLCM and GLSZM
features followed by GLDM and NGTDM were the most
selected features, respectively, showing the highest correla-
tion of the texture, hence the heterogeneity of the underly-
ing biology of the myocardial tissue with the outcome of
interest. First-order radiomic features and GLRLM in com-
bined features were not selected by RFE. In addition, in both
cases, only one feature in shape family was selected. Since
the region of interest was set as the whole left ventricle, it
was expected that morphological features do not show much
correlation with the desired outcome.
Toward the identification of the optimum automated
model, different combinations of RFE feature selection and
machine learning algorithms were applied on the included
feature sets (ConQuaFea, radiomics, and combined). Con-
QuaFea feature sets showed superiority over radiomic fea-
tures. Nevertheless, radiomic features also showed an ade-
quate performance. However, it was suggested that further
studies should investigate the correlations between different
feature sets in an independent analysis.
In summary, our study highlighted the potential of MPI
SPECT ConQuaFea. Radiomic features for the identification
of left ventricular contractile patterns also showed an accept-
able performance. Physicians may benefit from evaluating
each patient’s specific condition and determine if CRT is
needed. Moreover, it can reduce the workload and time spent
by physicians, since the detection of U-shaped contractile
patterns takes an average of 5min per patient. In addition,
excessive workloads, long working hours, sleep depriva-
tion, expertise of the physician, and other factors can lead
to misdiagnosis. Hence, this study attempts to help physi-
cians making better and more precise decisions in short time.
As mentioned earlier, despite the effectiveness and
importance of CRT [6, 7], at least 30% of patients selected
for this treatment do not respond well to this costly and
invasive treatment [8–12], showing insufficiency of current
inclusion criteria for this treatment. Moreover, from the 29
patients treated with CRT, 13 patients (45%) did not respond
to the treatment (Table4). A number of studies attempted
to introduce new criteria to improve the response rate to
treatment. In Bax etal. [14, 47], the presence of left ven-
tricular mechanical dyssynchrony was more common in
patients responding to CRT treatment, and therefore, it was
suggested as a criterion to improve response to CRT treat-
ment. A study by Adelstein etal. [9] also suggested that CRT
leads should not be placed in areas of scar tissue diagnosed
by MPI prior to implantation, as otherwise the response to
treatment will be reduced. In the study of Chen etal. [15],
the authors realized that the only characteristics which were
different between the CRT patients who were responders
Journal of Digital Imaging
1 3
and those who were not, were the histogram bandwidth and
phase standard deviation (extracted from phase analysis
using Emory Cardiac Toolbox at the baseline (before CRT)),
which is related to left ventricular mechanical dyssynchrony.
He etal. [16] conducted research with the goal of using deep
learning to extract new LVMD features using GSPECT MPI
phase analysis to help selecting CRT patients. New LVMD
parameters retrieved by automatedAutoencoder (AE) from
GSPECT MPI have the potential to enhance response pre-
diction prior to CRT. The study by Rastgou etal. [48] stated
that the phase analysis parameters of Emory Cardiac Tool-
box and QGS program are well correlated, but these param-
eters should not be used interchangeably. Furthermore, in
relation to entropy (a parameter in phase analysis), the lower
value of this parameter indicates synchronized heart contrac-
tion whereas the higher value indicates desynchronize heart
contraction [49].
In our study, the QGS program was used to extract phase
analysis indices, and when the features were selected by
RFE, the apex standard deviation and inferior standard
deviation from ConQuaFea data set and apex entropy and
lateral bandwidth from combined features were selected.
Since these features are related to the left ventricular con-
tractile pattern and the left ventricular contractile pattern to
the type of CRT response, the phase analysis parameters,
as mentioned in the above studies, are related to the type of
CRT response.
In a study by Feeny etal. [50], a machine learning model
was developed for CRT outcome prediction by enrolling 925
patients, with 9 clinical features (QRS morphology, QRS
duration, New York Heart Association classification, left
ventricular ejection fraction and end-diastolic diameter, sex,
ischemic cardiomyopathy, atrial fibrillation, and epicardial
left ventricular lead) used as input for the Naïve Bayes clas-
sifier. Their machine learning model outperformed the con-
ventional guideline with an increased AUC, 0.70 vs. 0.65
(p value < 0.02), and an increased event-free survival with
concordance index = 0.61 vs. 0.56 (p value < 0.001).
In a study by Tao etal. [19], contractile patterns were
used to analyze CRT responses. Their results showed that
89% of U-shaped group were responders to CRT and 11%
were non-responders. In our study, of the 29 patients who
performed CRT, 16 patients had U-shaped contractile pat-
terns, from which 11 responded positively, and 13 had
non-U-shaped contractile patterns, among which 11 did
not respond to the treatment (Table4). It led to a success
rate of 76% compared to traditional patient selection (16
responders out of 29 (55% responders)) in predicting treat-
ment outcome based on contractile patterns. We also ana-
lyzed the treatment response, in case CRT was prescribed
based on the left ventricle contractile pattern detected by
our proposed models (Table4). The results were slightly
lower than the model based on the ground truth status of
the left ventricle contractile pattern identified by the two
nuclear medicine physicians which was expected since the
error of pattern identification is also introduced to the final
results. Moreover, it should be noted that according to the
study conducted by Hartlage etal. [17], LV lead concord-
ant to the latest contracting site would be more likely to
produce a superior CRT response beside the left ventricu-
lar contractile pattern. In fact, in their study, patients with
a U-shaped contraction pattern and the LV lead concordant
to the last contraction site were 92% respondent, which
indicates the importance of examining the pattern as well
as the correct location of the lead in CRT patients. In our
study, the main purpose was to detect the left ventricular
contractile pattern, and therefore, the correct location of
the leads was not investigated.
This study inherently bears a few limitations. Regard-
ing the task of CRT outcome prediction, we set the criteria
on the basis of the left ventricular contractile pattern and
did not consider matters, such as the location of the CRT
leads. Future studies might consider both to develop more
comprehensive automated models for this important task.
In addition, our dataset was obtained from one institute
which undermines the robustness of the findings. Further
studies might gather larger and diverse dataset obtained
from multiple institutes with different image acquisition
parameters and patients’ ethnicity to improve the reproduc-
ibility of the models [51]. However, for proof-of-concept,
this study contained enough patients.
Conclusion
In this study, machine learning models were developed
to predict left ventricular contractile patterns with Con-
QuaFea and radiomic features from GSPECT MPI using
different machine learning approaches with acceptable
and promising results. ConQuaFea performed better than
radiomic features in recognizing left ventricular contrac-
tile pattern. In addition, by diagnosing the patients’ left
ventricular contraction pattern, it is possible to improve
the patient’s selection for CRT treatment.
Supplementary Information The online version contains supplemen-
tary material available at https:// doi. org/ 10. 1007/ s10278- 022- 00705-9.
Funding Open access funding provided by University of Geneva.
This work was supported by Iran University of Medical Sciences and
Rajaie Cardiovascular Medical and Research Center under grant num-
ber IR.IUMS.FMD.REC.1400.087 and the Swiss National Science
Foundation under Grant SNRF 320030_176052.
Availability of Data and Material Not applicable.
Journal of Digital Imaging
1 3
Code Availability Open-source library including 3D-Slicer and Python
library used in this study.
Declarations
Ethics Approval This retrospective study was approved by the ethics
committee of Iran University of Medical Sciences (IR.IUMS.FMD.
REC.1400.087).
Consent to Participate Informed consent was waived by ethics groups.
Consent for Publication Informed consent was waived by ethics
groups.
Conflict of Interest The authors declare no competing interests.
Open Access This article is licensed under a Creative Commons Attri-
bution 4.0 International License, which permits use, sharing, adapta-
tion, distribution and reproduction in any medium or format, as long
as you give appropriate credit to the original author(s) and the source,
provide a link to the Creative Commons licence, and indicate if changes
were made. The images or other third party material in this article are
included in the article's Creative Commons licence, unless indicated
otherwise in a credit line to the material. If material is not included in
the article's Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder. To view a
copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.
References
1. Groenewegen A, Rutten FH, Mosterd A, Hoes AW: Epidemiology
of heart failure. Eur J Heart Fail 22:1342-1356, 2020
2. Nagueh SF: Mechanical dyssynchrony in congestive heart failure:
diagnostic and therapeutic implications. J Am Coll Cardiol 51:18-
22, 2008
3. Levy WC, etal.: The Seattle Heart Failure Model: prediction of
survival in heart failure. Circulation 113:1424-1433, 2006
4. Marín-Oyaga V, Gutiérrez-Villamil C, Dueñas-Criado K, Arévalo-
Leal S: Análisis de fase para la valoración del sincronismo
mecánico del ventrículo izquierdo mediante Gated-SPECT de
perfusión miocárdica. Importancia de los parámetros clínicos y
técnicos. Rev Fac Med 65:453–459, 2017
5. Yaghoobi N, Malek H: The age of reason for gated SPECT MPI
to deal with cardiac dyssynchrony. Res Cardiovasc Med 4, 2015
6. Bax JJ, etal.: Cardiac resynchronization therapy: part 1—issues
before device implantation. J Am Coll Cardiol 46:2153-2167, 2005
7. O'Brien T, Park M-S, Youn J-C, Chung ES: The past, present and
future of cardiac resynchronization therapy. Korean Circ J 49:384-
399, 2019
8. Abraham WT, Hayes DL: Cardiac resynchronization therapy for
heart failure. Circulation 108:2596-2603, 2003
9. Adelstein EC, Saba S: Scar burden by myocardial perfusion imaging
predicts echocardiographic response to cardiac resynchronization
therapy in ischemic cardiomyopathy. Am heart J 153:105-112, 2007
10. Auricchio A, etal.: Effect of pacing chamber and atrioventricular
delay on acute systolic function of paced patients with congestive
heart failure. Circulation 99:2993-3001, 1999
11. Leclercq C, Hare JM: Ventricular resynchronization: current state
of the art. Circulation 109:296-299, 2004
12. Leclercq C, Kass DA: Retiming the failing heart: principles and
current clinical status of cardiac resynchronization. J Am Coll
Cardiol 39:194-201, 2002
13. Glikson M, etal.: 2021 ESC Guidelines on cardiac pacing and
cardiac resynchronization therapy: Developed by the Task Force
on cardiac pacing and cardiac resynchronization therapy of the
European Society of Cardiology (ESC) With the special contri-
bution of the European Heart Rhythm Association (EHRA). Eur
Heart J. 42:3427-3520, 2021
14. Bax JJ, etal.: Left ventricular dyssynchrony predicts response
and prognosis after cardiac resynchronization therapy. J Am Coll
Cardiol 44:1834-1840, 2004
15. Chen J, etal.: Assessment of left ventricular mechanical dys-
synchrony by phase analysis of ECG-gated SPECT myocardial
perfusion imaging. J Nucl Cardiol. 15:127-136, 2008
16. He Z, etal.: A method using deep learning to discover new pre-
dictors of CRT response from mechanical dyssynchrony on gated
SPECT MPI. J Nucl Cardiol :210601355, 2021
17. Hartlage GR, etal.: Prediction of response to cardiac resynchro-
nization therapy using left ventricular pacing lead position and
cardiovascular magnetic resonance derived wall motion patterns:
a prospective cohort study. J Cardiovasc Magn Reson 17:1-10,
2015
18. Sohal M, etal.: Noninvasive assessment of LV contraction pat-
terns using CMR to identify responders to CRT. JACC Cardiovasc
Imaging 6:864-873, 2013
19. Tao N, etal.: Assessment of left ventricular contraction patterns
using gated SPECT MPI to predict cardiac resynchronization
therapy response. J Nucl Cardiol 25:2029-2038, 2018
20. Duckett SG, etal.: Relationship between endocardial activation
sequences defined by high-density mapping to early septal con-
traction (septal flash) in patients with left bundle branch block
undergoing cardiac resynchronization therapy. Europace 14:99-
106, 2012
21. Fung JW, etal.: Effect of left ventricular endocardial activation
pattern on echocardiographic and clinical response to cardiac
resynchronization therapy. Heart 93:432-437, 2007
22. Jackson T, etal.: A U-shaped type II contraction pattern in patients
with strict left bundle branch block predicts super-response to car-
diac resynchronization therapy. Heart rhythm 11:1790-1797, 2014
23. Boogers MM, Chen J, Bax JJ: Role of nuclear imaging in cardiac
resynchronization therapy. Expert Rev Cardiovasc Ther 7:65-72,
2009
24. Manafi-Farid R, etal.: [(18)F]FDG-PET/CT radiomics and arti-
ficial intelligence in lung cancer: Technical aspects and potential
clinical applications. Semin Nucl Med, 2022
25. Khodabakhshi Z, etal.: Overall Survival Prediction in Renal Cell
Carcinoma Patients Using Computed Tomography Radiomic and
Clinical Information. J Digit Imaging 34:1086-1098, 2021
26. Amini M, etal.: Overall Survival Prognostic Modelling of Non-
small Cell Lung Cancer Patients Using Positron Emission Tomog-
raphy/Computed Tomography Harmonised Radiomics Features:
The Quest for the Optimal Machine Learning Algorithm. Clin
Oncol (R Coll Radiol) 34:114-127, 2022
27. Amini M, etal.: Multi-level multi-modality (PET and CT) fusion
radiomics: prognostic modeling for non-small cell lung carci-
noma. Phys Med Biol 66, 2021
28. Khodabakhshi Z, Mostafaei S, Arabi H, Oveisi M, Shiri I, Zaidi
H: Non-small cell lung carcinoma histopathological subtype
phenotyping using high-dimensional multinomial multiclass CT
radiomics signature. Comput Biol Med 136:104752, 2021
29. Arian F, etal.: Myocardial Function Prediction After Coronary
Artery Bypass Grafting Using MRI Radiomics Features and
Machine Learning Algorithms. J Digit Imaging, 2022
Journal of Digital Imaging
1 3
30. Shayesteh S, etal.: Treatment response prediction using MRI-
based pre-, post-, and delta-radiomic features and machine learning
algorithms in colorectal cancer. Med Phys 48:3691-3701, 2021
31. Avard E, etal.: Non-contrast Cine Cardiac Magnetic Resonance
image radiomics features and machine learning algorithms for
myocardial infarction detection. Comput Biol Med 141:105145,
2022
32. Shiri I, etal.: Diagnosis of COVID-19 using CT image radiom-
ics features: a comprehensive machine learning study involving
26,307 patients. medRxiv, 2021
33. Shiri I, etal.: High-Dimensional Multinomial Multiclass Severity
Scoring of COVID-19 Pneumonia Using CT Radiomics Features
and Machine Learning Algorithms. Sci Rep:In press, 2022
34. Shiri I, etal.: Impact of feature harmonization on radiogenom-
ics analysis: Prediction of EGFR and KRAS mutations from
non-small cell lung cancer PET/CT images. Comput Biol Med
142:105230, 2022
35. Ashrafinia S, Dalaie P, Sadaghiani MS, Schindler T, Pomper MG,
Rahmim A: Radiomics analysis of clinical myocardial perfusion
stress SPECT images to identify coronary artery calcification.
medRxiv, 2021
36. Rahmim A, etal.: Improved prediction of outcome in Parkinson’s
disease using radiomics analysis of longitudinal DAT SPECT
images. NeuroImage: Clinical 16:539–544, 2017
37. Edalat-Javid M, etal.: Cardiac SPECT radiomic features repeat-
ability and reproducibility: A multi-scanner phantom study. J Nucl
Cardiol :1–15, 2020
38. Bi Q, Goodman KE, Kaminsky J, Lessler J: What is machine
learning? A primer for the epidemiologist. Am. J. Epidemiol.
188:2222-2239, 2019
39. Martin-Isla C, etal.: Image-based cardiac diagnosis with machine
learning: a review. Front Cardiovasc Med 7:1, 2020
40. Zwanenburg A, etal.: The image biomarker standardization ini-
tiative: standardized quantitative radiomics for high-throughput
image-based phenotyping. Radiology 295:328-338, 2020
41. Zaret BL, Beller GA: Clinical Nuclear Cardiology: State of the
Art and Future Directions E-Book, 2010
42. Garcia EV, etal.: Quantification of rotational thallium-201 myo-
cardial tomography. J. Nucl. Med. 26:17-26, 1985
43. Caldwell JH, Williams DL, Harp GD, Stratton J, Ritchie J: Quantita-
tion of size of relative myocardial perfusion defect by single-photon
emission computed tomography. Circulation 70:1048-1056, 1984
44. D. Cerqueira NJW, Vasken Dilsizian, Alice K. Jacobs, Sanjiv
Kaul, Warren K. Laskey: Standardized myocardial segmenta-
tion and nomenclature for tomographic imaging of the heart: a
statement for healthcare professionals from the Cardiac Imaging
Committee of the Council on Clinical Cardiology of the American
Heart Association. Circulation:539–542, 2002
45. Rubeaux M, Xu Y, Germano G, Berman DS, Slomka PJ: Normal
databases for the relative quantification of myocardial perfusion.
Curr. Cardiovasc. Imaging Rep 9:1-11, 2016
46. Van Griethuysen JJ, etal.: Computational radiomics system to
decode the radiographic phenotype. Cancer Res. 77:e104-e107, 2017
47. Bax JJ, etal.: Left ventricular dyssynchrony predicts benefit of
cardiac resynchronization therapy in patients with end-stage heart
failure before pacemaker implantation. Am. J. Cardiol. 92:1238-
1240, 2003
48. Rastgou F, etal.: Assessment of left ventricular mechanical dys-
synchrony by phase analysis of gated-SPECT myocardial perfu-
sion imaging and tissue Doppler imaging: comparison between
QGS and ECTb software packages. J. Nucl. Cardiol. 21:1062-
1071, 2014
49. Baughan N, Poitrasson-Rivière A, Moody JB, Lee BC, Ficaro EP:
Quantification of Cardiac Dyssynchrony Entropy Implemented
in a Commercial Nuclear Medicine Software Application. Nucl
206:977-983, 2020
50. Feeny AK, etal.: Machine learning prediction of response to
cardiac resynchronization therapy: improvement versus current
guidelines. Circ Arrhythm Electrophysiol 12:e007316, 2019
51. Shiri I, etal.: Decentralized Distributed Multi-institutional PET
Image Segmentation Using a Federated Deep Learning Frame-
work. Clin Nucl Med 47:606-617, 2022
52. M. M. Henneman etal.: Can LV dyssynchrony as assessed with
phase analysis on gated myocardial perfusion SPECT predict
response to CRT?, J. Nucl. Med., 48:1104–1111, 2007
53. H. Abu Daya, S. Malhotra, and P. Soman: Radionuclide Assess-
ment of Left Ventricular Dyssynchrony, Cardiol. Clin., 34:101–
118, 2016
54. C. Stellbrink, O. Breithardt, A. Sinha, P. Hanrath, M. K. I, and K.
D. R. Aachen: How to discriminate responders from non-respond-
ers to cardiac resynchronisation therapy, 6:101–105, 2004
Publisher's Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.