Informatics in Medicine Unlocked 32 (2022) 101055

Contents lists available at ScienceDirect

Informatics in Medicine Unlocked

journal homepage: www.elsevier.com/locate/imu

Deep learning applications in myocardial perfusion imaging, a systematic

review and meta-analysis
Ebraham Alskaf a, *, Utkarsh Dutta b, Cian M. Scannell a, Amedeo Chiribiri a
a
School of Biomedical Engineering & Imaging Sciences, King’s College London, United Kingdom
b
GKT, School of Medicine, King’s College London, United Kingdom

A R T I C L E I N F O A B S T R A C T

Keywords: Background: Coronary artery disease (CAD) is a leading cause of death worldwide, and the diagnostic process
Deep learning comprises of invasive testing with coronary angiography and non-invasive imaging, in addition to history,
Myocardial perfusion imaging clinical examination, and electrocardiography (ECG). A highly accurate assessment of CAD lies in perfusion
Cardiac magnetic resonance
imaging which is performed by myocardial perfusion scintigraphy (MPS) and magnetic resonance imaging (stress
Coronary artery disease
CMR). Recently deep learning has been increasingly applied on perfusion imaging for better understanding of the
diagnosis, safety, and outcome of CAD.
The aim of this review is to summarise the evidence behind deep learning applications in myocardial perfusion
imaging.
Methods: A systematic search was performed on MEDLINE and EMBASE databases, from database inception until
September 29, 2020. This included all clinical studies focusing on deep learning applications and myocardial
perfusion imaging, and excluded competition conference papers, simulation and animal studies, and studies
which used perfusion imaging as a variable with different focus. This was followed by review of abstracts and full
texts. A meta-analysis was performed on a subgroup of studies which looked at perfusion images classification. A
summary receiver-operating curve (SROC) was used to compare the performance of different models, and area
under the curve (AUC) was reported. Effect size, risk of bias and heterogeneity were tested.
Results: 46 studies in total were identified, the majority were MPS studies (76%). The most common neural
network was convolutional neural network (CNN) (41%). 13 studies (28%) looked at perfusion imaging classi­
fication using MPS, the pooled diagnostic accuracy showed AUC = 0.859. The summary receiver operating curve
(SROC) comparison showed superior performance of CNN (AUC = 0.894) compared to MLP (AUC = 0.848). The
funnel plot was asymmetrical, and the effect size was significantly different with p value < 0.001, indicating
small studies effect and possible publication bias. There was no significant heterogeneity amongst studies ac­
cording to Q test (p = 0.2184).
Conclusion: Deep learning has shown promise to improve myocardial perfusion imaging diagnostic accuracy,
prediction of patients’ events and safety. More research is required in clinical applications, to achieve better care
for patients with known or suspected CAD.

1. Introduction management of patients with CAD has significant impact on both

morbidity and mortality.
1.1. Background There have been significant advancements in CAD imaging in the last
two decades, from anatomical imaging of the coronary tree by means of
Coronary artery disease (CAD) continues to be a major cause of death invasive x-ray coronary angiography and cardiac computed tomography
and hospitalisation worldwide including in high-income countries [1]. (CCTA), to functional assessment of coronary stenoses and their impact
The main underlying pathology lies in the progressive nature of coro­ on the myocardium both at rest and stress (physical or pharmacolog­
nary atherosclerotic process. Therefore, timely diagnosis to aid ical), using stress echocardiography, nuclear myocardial perfusion

* Corresponding author. School of Biomedical Engineering & Imaging Sciences, King’s College London, 4th Floor Lambeth Wing, St Thomas’ Hospital, Westminster
Bridge Road, London, SE1 7EH, United Kingdom..
E-mail address: [email protected] (E. Alskaf).

https://doi.org/10.1016/j.imu.2022.101055
Received 27 June 2022; Received in revised form 10 August 2022; Accepted 12 August 2022
Available online 18 August 2022
2352-9148/Crown Copyright © 2022 Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
E. Alskaf et al. Informatics in Medicine Unlocked 32 (2022) 101055

scanning (MPS), and stress perfusion cardiac magnetic resonance process. The analysis focused on participants with known or suspected
(CMR). Myocardial perfusion abnormalities are one of the early stages in CAD who had a perfusion imaging modality with the application of deep
the ischaemic cascade and ischaemic constellation, which also includes learning. Comparison was made with the standard imaging tests used in
angina symptoms, electrocardiographic (ECG) changes and ventricular clinical practice to identify the functional significance of coronary artery
wall motion abnormalities [2]. lesions (index test). A clinical reference standard is used for both tech­
Another exciting advancement has been made in computer vision niques (reference test) which is considered the gold standard.
technology following the revolution of neural networks and artificial Medical imaging techniques presented in conferences as part of
intelligence (AI) algorithms. Deep learning is the main subfield of AI challenges, such as Medical Image Computing and Computer Assisted
which has been the focus of computing in medical imaging, with car­ Intervention (MICCAI), simulation studies and animal studies were not
diovascular imaging being one of the common arenas for such novel included due to the ambiguity in their direct relation to patient care.
applications. Cardiac perfusion imaging is one of the main applications Given that the main scope of this review is on the direct application of
which has been studied by many deep learning practitioners and com­ deep learning on myocardial perfusion imaging, studies which used
puter vision experts. perfusion data as an input variable for prediction without deep learning
One of the key aspects of deep learning is that it allows automation of image applications were excluded. As there are numerous studies of left
clinical tasks, and thus reduces dependence on users. This has significant ventricular segmentation using deep learning, these studies were not
advantage in perfusion imaging interpretation given that the diagnostic included unless they form the basis for perfusion quantification. Finally,
accuracy of visual assessment by users is highly dependent on level of studies of automated perfusion quantification which relied mainly on
training, and previously it has been demonstrated that automated hand crafted algorithms or non-deep learning algorithms such as prin­
quantitative analysis performed similar to highly trained users (level 3) ciple component analysis (PCA) were not included.
in interpreting perfusion CMR imaging [3].
2.3. Search procedure
1.2. Rationale and objectives
For published literature, we searched MEDLINE (with PubMed
There is mounting evidence of the successful applications of deep extension) and EMBASE using Ovid search engine. To include all
learning in cardiac perfusion imaging, as demonstrated by the increasing possible Medline Subject Headings [MeSH] terms, Yale Mesh Analyzer
number of publications. Moreover, data derived from medical imaging was used after identifying two studies manually on MEDLINE database
can be integrated into specific machine learning approaches to provide with a focus on deep learning and CAD perfusion imaging modalities.
valuable information for the prediction of different outcomes by The PMIDs for those papers were inserted into the analyser and the
exploring new correlations between variables and clinical data to build resulting MeSH terms were used as a guide in the systematic search.
predictive models. Truncation was used in imaging terms [imag*], cardiac [cardia*],
As a result, it is becoming increasingly important that the current myocardial [myocardia*], [quantif*] and coronary [coronar*]. Wild­
literature and evidence behind deep learning applications in myocardial cards were used with one term [isch?emi*]. Plain terms were used for
perfusion imaging needs further evaluation, as well as recommendations [‘perfusion’], [‘stress’], [‘deep learning’], [‘machine learning’], [‘neural
of how to fine-tune the research towards more meaningful results for networks’], [‘artificial intelligence’], [‘supervised learning’], [‘unsu­
patients. pervised learning’], and [‘semi-supervised learning’]. The search
Therefore, the objective of this review is to determine the diagnostic included all records from database inception until September 29, 2020
accuracy of cardiac perfusion imaging using deep learning algorithms, with no language constraints. Full Ovid search strategy and output is
the impact of deep learning on image quality, image safety, and the shown in Appendix [1]. No routine use of methodology search filters has
assessment of its prognostic value. been used due to reports of missing relevant studies and inconsistency
[4].
2. Methods To avoid publication bias and give currency to this systematic review
with upcoming research, the grey literature also has been searched. This
2.1. Design includes:

The umbrella protocol for this systematic review is registered in the • Web of Science Conference Proceedings.
International Prospective Register of Systematic Reviews (PROSPERO, • Open Grey database.
CRD42020204164), and reported according to the Preferred Reporting • Manual searching of references
Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.
This review follows the Cochrane Review structure of Diagnostic Test 2.4. Data extraction
Accuracy (DTA) [4]. All searching activities were performed by two
independent authors (EA and UD), with divergences solved after The extracted summary estimates included imaging modality per­
consensus. formance after the application of deep learning (sensitivity, specificity,
The main review question was determined using the PICO approach: and area under the curve (AUC)). The sample size of each study with the
imaging modality used and deep learning techniques were all reported.
• Population: patients with suspected or known coronary artery dis­ The following is a summary of input data which were reported from
ease (CAD) each study in the review:
• Intervention: deep learning applications in CAD perfusion imaging
• Comparison: comparison with conventional CAD imaging 1. First author’s surname
• Outcome: improve test accuracy and patient care 2. Year of publication
3. Model output
2.2. Selection criteria 4. Total number of participants
5. Imaging modality used
Selection criteria decision was made by one author (EA) and over- 6. Index test
read by a senior author (AC), with disagreement resolved after 7. Reference test
consensus. Both prospective and retrospective studies were included 8. Deep learning methods
with no restrictions based on minimal sample sizes or recruitment 9. Sensitivity

2
E. Alskaf et al. Informatics in Medicine Unlocked 32 (2022) 101055

Fig. 1. PRISMA flow diagram showing the systematic search strategy.

10. Specificity effects meta-analysis with logit-transformed pairs of sensitivities and

11. AUC false positive rates method. Two SROC plots were performed using
linear mixed model to compare convolutional neural network (CNN)
performance against multi-layer perceptron (MLP). Publication bias and
2.5. Statistical analysis effect size derived from each study accuracy compared to mean accuracy
was tested using funnel plot and Egger’s test. P value of less than 0.05
The diagnostic accuracy of the imaging modalities was measured was considered significant. Heterogeneity was examined using tau2, I2
mainly with specificity and sensitivity analyses and presented as forest and Q tests.
plots. Data were reported as count or percentages. All statistical analysis was performed using RStudio software Version
Given that most studies did not report the values for true positive April 1, 1106 using R programming language version 4.0.4, “mada” and
(TP), false positive (FP), false negative (FN), and true negative (TN), a “meta” packages were used for meta-analysis.
confusion matrix was generated for each of the included studies in the
meta-analysis by taking sample size (S) to calculate FN using sensitivity 3. Results
and FP using specificity. This was followed by subtracting FN from S to
calculate TN and FP from S to calculate TP. 3.1. Search results
Although different studies reported different perfusion interpreta­
tion scales in MPS imaging, there were 2 common scales: a binary scale 715 study entries from the published literature Ovid search, and 432
of normal vs abnormal, and an ordinal scale from 0 to 4. Both scaling entries from grey literature were identified. After the screening of titles
methods are considered similar given that the ordinal scale would group and duplicate selection, 320 studies were included in the initial analysis
0 and 1 as normal, and 2,3 and 4 as abnormal. The ground truth finding for which titles indicated that the study might be of relevance. Following
from the reference test was considered the threshold for the summary full text review, 46 studies were included in the systematic review, of
receiver-operating curve (SROC) with bivariate diagnostic random

3
E. Alskaf et al. Informatics in Medicine Unlocked 32 (2022) 101055

Table 1
List of all relevant studies included in this systematic review.
First author Year Model output Sample size Imaging Model Index test Reference test External
modality validationa

Fujita et al. 1992 Perfusion classification (8 74 MPS MLP Expert reader Invasive coronary No
[5] classes) angiography
Wang et al. 1993 Perfusion classification (2 100 MPS MLP No Invasive coronary No
[6] classes x 64 segments) angiography
Porenta et al. 1994 Perfusion classification (2 159 MPS MLP Expert reader Invasive coronary No
[7] classes x 2 segments) angiography (81
cases)
Hamilton 1995 Perfusion classification (2 410 MPS MLP No Expert reader No
et al. [8] classes x 24 segments)
Goodenday 1997 Perfusion classification (2 × 42 MPS MLP Expert reader Invasive coronary No
et al. [9] 3 classes) angiography
Lindahl et al. 1997 Perfusion classification (2 135 MPS MLP Expert reader Invasive coronary No
[10] classes x 2 regions) angiography
Scott et al. 2004 Coronary artery disease 102 MPS MLP Expert reader Invasive coronary No
[11] prediction angiography
Ohlsson et al. 2004 Perfusion classification (5 1320 MPS MLP No Expert reader No
[12] classes x 5 segments)
Tagil et al. 2008 Perfusion classification (2 316 MPS MLP, KNN Logistic Regression Expert reader No
[13] classes x 5 segments)
Lomsky et al. 2008 Perfusion classification (2 950 MPS MLP Emory cardiac toolbox Expert reader Yes
[14] classes x 5 segments)
Guner et al. 2010 Perfusion classification (5 308 MPS MLP Expert reader Invasive coronary No
[15] classes) angiography
Abbasi et al. 2012 Perfusion classification (2 208 MPS MLP No Expert reader No
[16] classes x 20 segments)
Arsanjani 2013 Perfusion classification (5 957 MPS SVM Expert reader Invasive coronary No
et al. [17] classes x 17 segments) angiography
Nakajima 2015 Perfusion classification (5 1157 MPS MLP Expert reader Invasive coronary Yes
et al. [18] classes x 17 segments) angiography
Xiong et al. 2015 Perfusion classification (2 140 CCTA AdaBoost Random Forest & Naïve- Invasive coronary No
[19] classes x 17 segments) Bayes angiography
Parages et al. 2016 Perfusion classification (5 280 MPS Naïve-Bayes Non-prewhitening Expert reader na
[20] classes x 17 segments) simulated,
133 clinical
Lee et al. [21] 2016 Ischaemia prediction from 250 CCTA Gradient FFRCT alone Invasive FFR No
FFRCT and rCTP Boost
Li et al. [22] 2017 Fully automated perfusion 21 MCE CNN + Active Shape Model Expert reader No
segmentation Random
Forest
Kim et al. [23] 2017 Automated perfusion 59 CMR Random Histogram of Oriented Expert reader No
landmarks detection (RV Forest Gradients
insertion point and LV centre
point)
Nakajima 2017 Perfusion classification (3 1365 MPS MLP No Expert reader No
et al. [24] classes x 17 segments)
Al Mallah 2017 Prediction of cardiac death 9026 MPS Random Logistic Regression na No
et al. [25] at median follow-up of 4.3 Forest
years
Nakajima 2018 Perfusion classification (5 1157 MPS MLP No Expert reader Yes
et al. [26] classes x 17 segments)
Do et al. [27] 2018 Fully automated perfusion 28 CMR CNN No Expert reader Yes
segmentation
Eisenberg 2018 Perfusion classification (5 1925 MPS LogitBoost Expert reader Invasive coronary No
et al. [28] classes x 17 segments) angiography
Betancur et al. 2019 Perfusion classification (2 1160 MPS CNN Automated cTPD Invasive coronary Yes (used leave-
[29] classes x 3 segments) angiography one-center-out
cross-validation)
Kim et al. [30] 2019 Fully automated perfusion 145 CMR CNN Semi-automated Expert reader No
quantification
Scannell et al. 2019 LV peak signal 175 CMR CNN No Expert reader No
[31] enhancement, LV bounding
box and segmentation, RV
insertion point
Spier et al. 2019 Perfusion classification (2 946 MPS CNN No Expert reader No
[32] classes x 17 segments)
Fan et al. [33] 2019 Accelerated k-space 40 CMR CNN Compressed sensing Expert reader No
perfusion processing reconstruction
Ko et al. [34] 2019 Perfusion attenuation map 502 MPS CNN No CT-based No
generation attenuation maps
Chiu et al. 2019 Perfusion classification (5 150 MPS CNN Emory cardiac toolbox Invasive coronary No
[35] classes x 17 segments) angiography
Song et al. 2019 119 MPS CNN Full dose No
[36] perfusion image
(continued on next page)

4
E. Alskaf et al. Informatics in Medicine Unlocked 32 (2022) 101055

Table 1 (continued )
First author Year Model output Sample size Imaging Model Index test Reference test External
modality validationa

Prediction of full dose Spatiotemporal non-local

perfusion image from means, Gaussian, Maximum-
reduced dose Likelihood
Rahmani et al. 2019 Coronary angiography 93 MPS MLP No Invasive coronary na
[37] outcome prediction (2 angiography
classes x 20 segments)
Singh et al. 2020 Prediction of MACE at 1185 MPS CNN Clinical model, ventricular na No
[38] median follow-up of 385 function model, absolute
days perfusion quantification
model, integrated model
Knott et al. 2020 Prediction of MACE and 1049 CMR CNN na na na
[39] death at median follow-up of
605 days
Shiri et al. 2020 Prediction of full time from 363 MPS CNN na Full time/ No
[40] half time perfusion projection
acquisition acquisition
perfusion
Ramon et al. 2020 Prediction of full dose 1052 MPS CNN na Full dose No
[41] perfusion image from perfusion image
reduced dose
Aposto- 2020 Perfusion classification (2 216 MPS CNN Expert reader Invasive coronary No
lopoulos classes x 17 segments) angiography
et al. [42]
Xue et al. [43] 2020 Automated perfusion 15,789 CMR CNN No Expert reader No
landmarks detection (LV
blood pool enhancement)
Berkaya et al. 2020 Perfusion quantification (3 192 MPS CNN No Expert reader No
[44] classes x 5 sections)
Shi et al. [45] 2020 Perfusion attenuation map 65 MPS CNN na CT-based No
generation attenuation maps
Hu et al. [46] 2020 Revascularisation prediction 1980 MPS LogitBoost Expert reader Invasive coronary No
per patient/per vessel from angiography
49 variables
Juarez- 2020 Prediction of MI and death at 951 MPS CNN (Cox- na Expert reader No
Orozco medial follow-up of 6 years Nnet)
et al. [47]
Shu et al. [48] 2020 Prediction of myocardial 154 CCTA SVM CTCA stenosis and radiomics Expert reader Yes
ischaemia of MPS signature
Cantoni et al. 2020 Prediction of CAD from 14 517 MPS Random MPS (C-SPECT) Expert reader No
[49] variables (CZT- Forest
SPECT)
Wang et al. 2020 Prediction of CAD from 5 88 MPS SVM 6 ML models (LDA, DT, KNN, Invasive coronary No
[50] variables LR, NB, RF) angiography

CMR, cardiac magnetic resonance; CNN, convolutional neural network, CT; computed tomography; CCTA, coronary CT angiography; DT, decision tree; FFR, fractional
flow reserve; KNN, K-nearest neighbours; LDA, latent Dirichlet algorithm; LR, logistic regression; LV, left ventricle; MACE, major adverse cardiovascular events; MCE,
myocardial contrast echocardiography; MI, myocardial infarction; MLP, multi-layer perceptron; MPS, myocardial perfusion scintigraphy; na, not applicable or not
available; NB, naïve-bayes; rCTP, resting CT perfusion; RF, random forest; RV, right ventricle; SPECT, single photon-emission computed tomography; SVM, support
vector machine.
a
External validation by using a completely separate dataset for testing or validation outside the original training dataset.

which 13 studies were included in the meta-analysis. The selection segment and classify perfusion imaging maps with various classes, most
procedure and results with reasons for exclusion in the full text assess­ of those studies were based on MPS imaging.
ment is illustrated in Fig. 1. A meta-analysis was performed on 13 studies where the output of the
classifier was based on perfusion maps segmentation and referenced to
3.2. Characteristics of studies the presence or absence of significant CAD based on invasive coronary
angiography or consensus of expert readers of MPS, a summary of their
The final number of studies included in this systematic review was corresponding sensitivity and specificity is depicted in the coupled forest
46, details of first author, year of publication, model output, sample size, plot Fig. 4. The plot shows good performance of the neural networks
machine learning and deep learning techniques, index test (compar­ with most studies reporting sensitivity and specificity of over 65%.
ator), and reference test (gold standard) are all given in Table 1. When comparing the performance of MLP with CNN across these
The majority of the studies were performed on MPS (76%). However, studies using the summary receiver operating curve (SROC), CNN
the number of studies in CMR has increased in the last 2 years, as shown showed a higher value of SROC (higher sensitivity, lower false positive
in Fig. 2. rate) with area under the curve (AUC) of 0.894, compared to MLP (AUC
The most common neural network architecture in early years was = 0.848), as showed in Fig. 5. The overall pooled AUC including all 13
MLP (35%), which has been dominated by CNN (41%) in recent years, as studies was averaged at 0.859, showing good performance.
shown in Fig. 3.
3.4. Assessment of heterogeneity
3.3. Meta-analysis of perfusion classification
Quantifying heterogeneity showed τ2 = 0.0037 with confidence in­
There were several studies which applied deep learning directly to terval [0.0000, 0.0295], which contains zero, indicating no significant

5
E. Alskaf et al. Informatics in Medicine Unlocked 32 (2022) 101055

Fig. 2. A stacked barplot showing the number of publications for each imaging modality in myocardial perfusion imaging over the last 30 years.
CCTA, coronary computed tomography angiography; CMR, cardiac magnetic resonance; MCE, myocardial contrast echocardiography; MPS; myocardial perfusion
scintigraphy.

between-study heterogeneity exists in our data. Taking all the above into consideration, a table of the included
I2 was found to be 22.3%, meaning that less than quarter of the studies with their associated risk of bias is shown in Appendix [2]. There
variation in our data is estimated to stem from true effect size differ­ were 13 studies (28%) which did not include an index test to compare
ences. Using literature “rule of thumb”, we can characterize this amount with the machine learning or deep learning model before comparing
of heterogeneity as mild. with the reference test (ground truth). All studies defined a ground truth
Predictive interval was found to be ranging from [0.8569 to 1.1645], test against which they tested the model performance, and the majority
meaning that it is possible that some future studies will likely find of the studies blinded the model reporters from the ground truth results.
positive effect based on the present evidence. This indicates the high reliability of the reported results.
Finally, the reported p value for Q test was found to be above sig­ Funnel plot in Fig. 6 shows asymmetrical pattern indicating small-
nificance level (p = 0.2184), meaning there is no significant study effects. Egger’s test was significant with p value < 0.001, in­
heterogeneity. dicates that the data in the funnel plot are indeed asymmetrical, and
possibly related to publication bias.
3.5. Assessment of risk of bias
4. Discussion
The risk of bias was assessed using the Quality Assessment of Diag­
nostic Accuracy Studies (QUADAS) tool. A modified version was 4.1. Deep learning techniques
adapted and five main fields were assessed:
The application of deep learning models on myocardial perfusion
1. Patient selection: a high quality study would randomly select pa­ imaging started in the early 1990s, where all the studies were focused on
tients from a population meeting the inclusion criteria. the use of MLP architecture and applied on MPS [5–10]. MLPs are
2. Index test: a high quality diagnostic test study would include a composed of three types of layers: input layers taking the raw image
comparator test. data, hidden layers which are connected via weight vectors, and an
3. Reference test: all diagnostic test studies should have a gold standard output layer which takes the weighted sum, applies an output function
test for validation. and returns a prediction [51]. The main output prediction of interest in
4. Index test results blinded: a high quality study would blind the re­ early studies was perfusion map classification, this has continued in
sults of the comparator test to the deep learning arm. early 2000s, when the performance of MLP was also compared to other
5. Reference test results blinded: a high quality study would blind the traditional linear and non-linear machine learning algorithms, such as
results of the gold standard test to the deep learning arm. K-nearest neighbours (KNN) [13] and support vector machines (SVM)
[17]. Most of the networks achieved high performance metrics when

6
E. Alskaf et al. Informatics in Medicine Unlocked 32 (2022) 101055

Fig. 3. A stacked barplot showing the number of studies for each machine and deep learning algorithm used in myocardial perfusion studies over the last 30 years.
CNN, convolutional neural network; MLP, multi-layer perceptron; SVM, support vector machine.

Fig. 4. Forest plot of both specificity and sensitivity reported by the 13 MPS studies looking at deep learning in perfusion images classification.

compared to human expert readers. significant limitation on the size of the model available to learn image
There has been a substantial increase in the number of publications features. The CNN overcomes this challenge by using convolutional
on deep learning in general with more focus on CNN over the last few layers which have significantly fewer parameters and make use of
years, as shown in Fig. 3. Due to the high dimensionality of imaging extensive weight sharing. The process of several convolutional layers
data, the fully connected layers which MLPs are based on put a can be thought in the following steps: detect low level features and edges

7
E. Alskaf et al. Informatics in Medicine Unlocked 32 (2022) 101055

stage of perfusion quantification in CMR by automated identification

of anatomical landmarks, such as the right ventricle (RV) insertion point
into the septum and left ventricle (LV) centre on peak contrast
enhancement [23,31,43]. Furthermore, CNN algorithms have been
successfully applied to the segmentation of CMR perfusion images [27,
30] with high performance. These applications in CMR still require
further research. Another exciting application of CNN is on k-space ac­
celeration and reconstruction for faster perfusion images acquisition
[33] which is another attractive research application for deep learning.
Another novel application on the horizon is using deep learning for
predicting myocardial blood flow in perfusion CMR using
physics-informed neural networks (PINNs) [52,53].
Other successful applications of deep learning include prediction,
whether looking at prediction of death or myocardial infarction [39,47],
prediction of revascularisation events [46], or prediction of high-quality
full radiation dose MPS images from low dose or short time scans which
has significant impact on the radiation dose delivered to the patients
[36,40,41], and the identification of acquisition sequence type and
image plane in CMR [54].
Fig. 5. Summary receiver operating curve (SROC) comparing between con­ Furthermore, there are newer deep learning techniques using
volutional neural network (CNN) and multi-layer perceptron (MLP). This shows generative adversarial network (GAN) which have some promising
a higher performance of CNN (solid line) compared to MLP (dotted line).
application for image reconstruction, but this is still an active area of
research.

4.3. Applicability of findings to review question

Given the evidence of multiple successful applications of deep

learning on perfusion imaging presented in this review, the value of this
evidence, although significant, remains in research applications with
limited clinical use. A wider use of such applications based on the evi­
dence presented could have significant impact on patients with known
or suspected CAD. Reducing scan time, radiation dose, human resources
and increasing diagnostic accuracy can save patients time, and result in
better management of their coronary artery disease which has signifi­
cant mortality and morbidity benefits.

5. Limitations
Fig. 6. Funnel plot showing asymmetry of the studies and significant variation
in their effect size values (p < 0.01). There are more applications and techniques which have been used
without full publications of clinical studies and were not reported in this
from raw pixel data in the early layers, use these edges to detect shapes review, given that the main scope is clinical applications of deep
in the later layers, and use these shapes to detect higher-level features learning in perfusion imaging.
for prediction. An additional useful property of CNNs is that they lend The published articles included in this review did not report the same
themselves well to be used with transfer learning where the majority of performance metrics, which was a challenge on the meta-analysis pro­
the network is kept with its high-level feature extraction ability and only cess, one of the main observations was that the performance metrics
the last output layer is exchanged with a new layer to fit with the pur­ were reported in some studies for both stress and rest images, but not in
pose of the study [51]. As a result, the majority of deep learning studies others. As a result, only the highest performance score of the models for
on perfusion imaging in the last few years have used CNNs as the main the stress images was reported in this meta-analysis.
architecture, as shown in Fig. 3. Furthermore, the power and flexibility
of CNNs has opened the window for deep learning applications in more 6. Conclusion
challenging image analysis domains such as stress perfusion CMR [23,
27,30,31], resting CT perfusion (rCTP), and myocardial contrast echo­ 6.1. Implications for practice
cardiography (MCE) [22].
In this review the evidence of successful deep learning applications
4.2. Summary of main results in myocardial perfusion imaging has been presented. Most of the early
studies used the standard MLP perceptron architecture on MPS imaging,
The performance of neural networks for the identification of perfu­ but more recently CNN architectures gained in popularity given its su­
sion defects has proven to have a comparative performance to human perior performance in image analysis, and deep learning applications
expert reading, and had a strong overall accuracy in MPS studies (AUC have expanded to other perfusion imaging modalities, mainly stress
0.859) regardless of the comparator or reference tests. The meta- perfusion CMR. The accuracy of deep learning has proven to be high in
analysis presented in this review also shows the superior performance perfusion image classification to diagnose CAD compared to human
of CNNs compared to MLPs in reading and classification of perfusion readers and conventional diagnostic procedures performed in routine
maps. clinical practice, based on our meta-analysis of the relevant studies.
The applications of deep learning on stress perfusion CMR has been
increasing in recent years. There are some promising data on the
effectiveness of using deep learning with CNNs to the pre-processing

8
E. Alskaf et al. Informatics in Medicine Unlocked 32 (2022) 101055

6.2. Implications for research [8] Hamilton D, Riley PJ, Miola UJ, Amro AA. A feed forward neural network for
classification of bull’s-eye myocardial perfusion images. Br J Radiol 1995 Nov;68
(815):1208–11.
The successful preliminary applications of deep learning in stress [9] Goodenday LS, Cios KJ, Shin I. Identifying coronary stenosis using an image-
perfusion CMR have opened a wide spectrum of potential applications to recognition neural network. IEEE Eng Med Biol Mag Sep-Oct 1997;16(5):139–44.
improve accuracy, accelerate scan times, and predict outcomes. Despite [10] Lindahl D, Palmer J, Ohlsson M, Peterson C, Lundin A, Edenbrandt L. Automated
interpretation of myocardial SPECT perfusion images using artificial neural
the high performance of deep learning in MPS image classification, networks. J Nucl Med 1997;38(12):1870–5.
which have shown promise for more than two decades, there is still a [11] Scott James A, Aziz Kusai, Yasuda Tsunehiro, Henry Gewirtz. Integration of clinical
lack of wide use in clinical practice. and imaging data to predict the presence of coronary artery disease with the use of
neural networks. Coron Artery Dis 2004 Nov;15(7):427–34.
As a result, the findings of this review would encourage more clinical [12] Ohlsson M. WeAidU - a decision support system for myocardial perfusion images
studies and trials to assess the performance and accuracy of deep using artificial neural networks. Artif Intell Med 2004;30(1):49–60.
learning in cardiac perfusion imaging using the latest techniques, in [13] Tägil K, Marving J, Lomsky M, Hesse B, Edenbrandt L. Use of neural networks to
improve quality control of interpretations in myocardial perfusion imaging. Int J
order to obtain clinical validation and to start to use this technology as Cardiovasc Imaging 2008;24(8):841–8.
clinical applications in perfusion imaging. Furthermore, other perfusion [14] Lomsky M, Gjertsson P, Johansson L, Richter J, Ohlsson M, Tout D, et al.
imaging modalities which are still in their infancy, such as rCTP and Evaluation of a decision support system for interpretation of myocardial perfusion
gated SPECT. Eur J Nucl Med Mol Imaging 2008;35(8):1523–9.
MCE, can also benefit from deep learning applications. [15] Guner LA, Karabacak NI, Akdemir OU, Karagoz PS, Kocaman SA, Cengel A, et al.
An open-source framework of neural networks for diagnosis of coronary artery
Funding disease from myocardial perfusion SPECT. J Nucl Cardiol 2010;17(3):405–13.
[16] Abbasi M, Farzanefar S, Emami A, Eftekhari M. Automated interpretation of
myocardial perfusion images with Multilayer Perceptron network: a decision
This research received no grant from any funding agency in the support system free from quantification and need for normal database. Eur J Nucl
public, commercial or not-for-profit sectors. This research has received Med Mol Imaging 2012;29(Suppl 2):S155–303.
[17] Arsanjani R, Xu Y, Dey D, Fish M, Dorbala S, Hayes S, et al. Improved accuracy of
grant from Wellcome Trust [222678/Z/21/Z]. myocardial perfusion SPECT for the detection of coronary artery disease using a
support vector machine algorithm. J Nucl Med 2013;54(4):549–55.
[18] Nakajima K, Matsuo S, Wakabayashi H, Yokoyama K, Bunko H, Okuda K, et al.
Authors contributions
Diagnostic performance of artificial neural network for detecting ischemia in
myocardial perfusion imaging. Circ J 2015;79(7):1549–56.
EA and UD have performed the systematic search, data extraction [19] Xiong G, Kola D, Heo R, Elmore K, Cho I, Min J. Myocardial perfusion analysis in
and writing the manuscript. EA and CS have contributed to data analysis cardiac computed tomography angiographic images at rest. Med Image Anal 2015
Aug;24(1):77–89.
and statistics. EA and AC have contributed to discussion and conclusion. [20] Parages F, O’Connor JM, Pretorius PH, Brankov JG. Machine-learning model
AC has contributed to final proof-reading as a senior author. All authors observer for detection and localization tasks in clinical SPECT-MPI. In: SPIE
have reviewed and approved the submission. proceedings, vol. 9787. Medical Imaging; 2016.
[21] Lee JH, Rizvi A, Han D, Hartaigh BO, Gransar H, Lin F, et al. Incremental benefit of
computed tomography derived resting myocardial perfusion using machine
learning method and computed tomography derived fractional flow reserve for
Declaration of competing interest predicting lesion ischemia. J Am Coll Cardiol 2016;67(13):1677.
[22] Li Y, Ho CP, Toulemonde M, Chahal N, Senior R, Tang MX. Fully automatic
myocardial segmentation of contrast echocardiography sequence using random
The authors declare that they have no known competing financial forests guided by shape model. IEEE Trans Med Imaging 2018;37(5):1081–91.
interests or personal relationships that could have appeared to influence [23] Kim YC, Chung Y, Choe YH. Automatic localization of anatomical landmarks in
the work reported in this paper. cardiac MR perfusion using random forests. Biomed Signal Process Control 2017;
38:370–8.
[24] Nakajima K, Kudo T, Nakata T, Kiso K, Kasai T, Taniguchi Y, et al. Diagnostic
Acknowledgements accuracy of an artificial neural network compared with statistical quantitation of
myocardial perfusion images: a Japanese multicenter study. Eur J Nucl Med Mol
Imaging 2017;44(13):2280–9.
None. [25] Al Mallah MH, Ahmed A, Elshawi R, Sakr S. Prediction of cardiac death after
pharmacological myocardial perfusion imaging using machine learning techniques.
Eur Heart J Cardiovasc Imaging Abstracts Supplement May 2017:i49–51.
Appendix A. Supplementary data
[26] Nakajima K, Okuda K, Watanabe S, Matsuo S, Kinuya S, Toth K, et al. Artificial
neural network retrained to detect myocardial ischemia using a Japanese
Supplementary data to this article can be found online at https://doi. multicenter database [Internet] Ann Nucl Med 2018;32(5):303–10. https://doi.
org/10.1016/j.imu.2022.101055. org/10.1007/s12149-018-1247-y. Available from:.
[27] Do HP, Yi Guo M, Yoon AJ, Nayak KS. Accuracy, uncertainty, and adaptability of
automatic myocardial ASL segmentation using deep CNN. arXiv 2018. https://doi.
References org/10.1002/mrm.28043 (December).
[28] Eisenberg E, Betancur J, Hu L, et al. The diagnostic accuracy of machine learning
from stress only fast-MPS. J Nucl Med May 1, 2018;59. supplement 1 508.
[1] Nowbar AN, Gitto M, Howard JP, Francis DP, Al-Lamee R. Mortality from ischemic
[29] Betancur J, Hu LH, Commandeur F, Sharir T, Einstein AJ, Fish MB, et al. Deep
heart disease: analysis of data from the world health organization and coronary
learning analysis of upright-supine high-efficiency SPECT myocardial perfusion
artery disease risk factors from NCD risk factor collaboration. Circ Cardiovasc Qual
imaging for prediction of obstructive coronary artery disease: a multicenter study.
Outcomes Jun 2019;12:e005375.
J Nucl Med 2019;60(5):664–70.
[2] Maznyczka A, Sen S, Cook C, Francis DP. The ischaemic constellation: an
[30] Kim YC, Kim KR, Choe YH. Automatic myocardial segmentation in dynamic
alternative to the ischaemic cascade—implications for the validation of new
contrast enhanced perfusion MRI using Monte Carlo dropout in an encoder-decoder
ischaemic tests. Open Hear. DOI: 10.1136/openhrt-2014-000178.
convolutional neural network. Comput Methods Programs Biomed 2020:185.
[3] Villa ADM, Corsinovi L, Ntalas I, Milidonis X, Scannell C, Di Giovine, et al.
https://doi.org/10.1016/j.cmpb.2019.105150.
Importance of operator training and rest perfusion on the diagnostic accuracy of
[31] Scannell CM, Veta M, Villa ADM, Sammut EC, Lee J, Breeuwer M, et al. Deep-
stressperfusion cardiovascular magnetic resonance. Journal of Cardiovascular
learning-based preprocessing for quantitative myocardial perfusion MRI. J Magn
Magnetic Resonance 2018;20(1):74. https://doi.org/10.1186/s12968-018-0493-4.
Reson Imaging 2020;51(6):1689–96.
[4] Leeflang MMG, Deeks JJ, Takwoingi Y, Macaskill P. Cochrane diagnostic test
[32] Spier N, Nekolla S, Rupprecht C, Mustafa M, Navab N, Baust M. Classification of
accuracy reviews. Systematic reviews 2013;16(149):889–97. https://doi.org/
polar maps from cardiac perfusion imaging with graph-convolutional neural
10.7326/0003-4819-149-12-200812160-00008.
networks. Sci Rep 2019;9(1):1–8.
[5] Fujita H, Katafuchi T, Uehara T, Nishimura T. Application of artificial neural
[33] Fan L, Shen D, Haji-Valizadeh H, Naresh NK, Carr JC, Freed BH, et al. Rapid
network to computer-aided diagnosis of coronary artery disease in myocardial
dealiasing of undersampled, non-Cartesian cardiac perfusion images using U-net.
SPECT bull’s-eye images. J Nucl Med 1992;33(2):272–6.
NMR Biomed 2020;33(5):1–8.
[6] Wang DC, Juni JE. Artificial neural network nuclear cardiac stress test
[34] Ko CL, Cheng MF, Yen RF, Chen CM, Lee WJ, Wang TD. Automatic alignment of
interpretation. IEEE Symposium on Nuclear Science (NSS/MIC) 1992:1282–4.
CZT myocardial perfusion SPECT and external non-contrast CT by deep-learning
[7] Porenta G, Dorffner G, Kundrat S, Petta P, Duit-Schedlmayer J, Sochor H.
model and dynamic data generation [Internet] J Nucl Med Conf 2019;60
Automated interpretation of planar thallium-201-dipyridamole stress-
(Supplement 1):500.
redistribution scintigrams using artificial neural networks. J Nucl Med 1994;35
(12):2041–7.

9
E. Alskaf et al. Informatics in Medicine Unlocked 32 (2022) 101055

[35] Chiu C-H, Cheng C-Y, Tseng T-W, Chen I-J, Yin Y-P. Deep learning analysis of Tl- [45] Shi L, Onofrey JA, Liu H, Liu YH, Liu C. Deep learning-based attenuation map
201 myocardial perfusion imaging for improve diagnostic accuracy. J Nucl Med generation for myocardial perfusion SPECT. Eur J Nucl Med Mol Imaging 2020;47
[Internet] 2019;60:201. (10):2383–95.
[36] Song C, Yang Y, Wernick MN, Pretorius PH, King MA. LOW-DOSE CARDIAC- [46] Hu LH, Betancur J, Sharir T, Einstein AJ, Bokhari S, Fish MB, et al. Machine
GATED SPECT studies using A residual convolutional neural network department learning predicts per-vessel early coronary revascularization after fast myocardial
of electrical and computer engineering. IL Department of Radiology , University of perfusion SPECT: results from multicentre REFINE SPECT registry. Eur Heart J
Massachussetts Medical School. Chicago: Illinois Institute of Technology; 2019. Cardiovasc Imaging 2020;21(5):549–59.
p. 653–6. Isbi. [47] Juarez-Orozco LE, Martinez-Manzanera O, van der Zant FM, Knol RJJ, Knuuti J.
[37] Singh A, Jerosch-Herold M, Bekele S, et al. Deep learning in quantitative PET Deep learning in quantitative pet myocardial perfusion imaging: a study on
myocardial perfusion imaging, a study on cardiovascular event prediction. JACC cardiovascular event prediction. JACC Cardiovasc Imaging 2020;13(1):180–2.
Cardiovasc Imaging 2020;13(1):167–82. [48] Shu ZY, Cui SJ, Zhang YQ, Xu YY, Hung SC, Fu LP, et al. Predicting chronic
[38] Rahmani R, Niazi P, Naseri M, Neishabouri M, Farzanefar S, Eftekhari M, et al. myocardial ischemia using CCTA-based radiomics machine learning nomogram
Improved diagnostic accuracy for myocardial perfusion imaging using artificial [Internet] J Nucl Cardiol 2020;(158). https://doi.org/10.1007/s12350-020-
neural networks on different input variables including clinical and quantification 02204-2. Available from:.
data. Rev Esp Med Nucl Imagen Mol 2019;38(5):275–9. [49] Cantoni V, Green R, Ricciardi C, Assante R, Zampella E, Nappi C, et al. A machine
[39] Knott KD, Seraphim A, Augusto JB, Xue H, Chacko L, Aung N, et al. The prognostic learning-based approach to directly compare the diagnostic accuracy of myocardial
significance of quantitative myocardial perfusion: an artificial intelligence-based perfusion imaging by conventional and cadmium-zinc telluride SPECT [Internet]
approach using perfusion mapping. Circulation 2020:1282–91. J Nucl Cardiol 2020. https://doi.org/10.1007/s12350-020-02187-0. Available
[40] Shiri I, AmirMozafari Sabet K, Arabi H, Pourkeshavarz M, Teimourian B, Ay MR, from:.
et al. Standard SPECT myocardial perfusion estimation from half-time acquisitions [50] Wang F, Xu W, Lv W, Du D, Feng H, Zhang X, et al. Evaluation of the diagnostic
using deep convolutional residual neural networks [Internet] J Nucl Cardiol 2020. value of joint PET myocardial perfusion and metabolic imaging for vascular
https://doi.org/10.1007/s12350-020-02119-y. Available from:. stenosis in patients with obstructive coronary artery disease [Internet] J Nucl
[41] Ramon AJ, Yang Y, Pretorius PH, Johnson KL, King MA, Wernick MN. Improving Cardiol 2020. https://doi.org/10.1007/s12350-020-02160-x. Available from:.
diagnostic accuracy in low-dose SPECT myocardial perfusion imaging with [51] Rosebrock A. Deep learning for computer vision 2019;1.
convolutional denoising networks. IEEE Trans Med Imaging 2020;39(9):2893–903. [52] van Herten RLM, Chiribiri A, Breeuwer M, Veta M, Scannell CM. Physics-informed
[42] Apostolopoulos ID, Papathanasiou ND, Spyridonidis T, Apostolopoulos DJ. neural networks for myocardial perfusion MRI quantification. Med Image Anal
Automatic characterization of myocardial perfusion imaging polar maps employing 2022;(78):102399. https://doi.org/10.1016/J.MEDIA.2022.102399.
deep learning and data augmentation. Hell J Nucl Med 2020;23(2):125–32. [53] Scannell CM, Bosch P vd, Chiribiri A, Lee J, Breeuwer M, Veta M. Deep learning-
[43] Xue H, Tseng E, Knott KD, Kotecha T, Brown L, Plein S, et al. Automated detection based prediction of kinetic parameters from myocardial perfusion MRI. 2019.
of left ventricle in arterial input function images for inline perfusion mapping using arXiv.org > eess > arXiv: 1907.11899.
deep learning: a study of 15,000 patients. Magn Reson Med 2020;84(5):2788–800. [54] Lim RP, Kachel S, Villa ADM, Kearney L, Bettencourt N, Young AA, et al. CardiSort:
[44] Kaplan Berkaya S, Ak Sivrikoz I, Gunal S. Classification models for SPECT a convolutional neural network for cross vendor automated sorting of cardiac MR
myocardial perfusion imaging. Comput Biol Med 2020;123(February). https://doi. images. 2021. arXiv Prepr. arXiv2109.08479.
org/10.1016/j.compbiomed.2020.103893.

10