Artificial Intelligence and Cardiac Imaging: We need to talk about this

Full text

154
Editorial
Artificial Intelligence and Cardiac Imaging: We need to talk
about this
José de Arimateia Batista Araujo-Filho1,2, Antonildes Nascimento Assunção Júnior1,3, Marco Antonio
Gutierrez3,Cesar Higa Nomura1,3
1Hospital Sirio Libanes, São Paulo, Brazil; 2Memorial Sloan-Kettering Cancer Center, New York, Estados Unidos; 3InCor-HFCMUSP, São Paulo, Brazil
Keywords
Artificial Intelligence; Machine Learning; Radiology.
Correspondence: José de Arimateia Batista Araujo-Filho •
E-mail: ariaraujocg@gmail.com
DOI: 10.5935/2318-8219.20190035
“Measure what is measurable, and make measurable what
is not so.”
Galileo Galilei (1564-1642)
With the rapid technological progress experienced by
medical imaging in recent years, the conversion of digital images
into high-dimensional data, that is, with a large number of
variables, has been driven by the concept that images contain
a myriad of underlying pathophysiological information that is
often difficult identify and comprehend using conventional
visual analysis.1 The quantitative analysis of these images and
the organization of these parameters in complex databases (Big
Data) — with large volume, variety and speed of information
generation — brought radiology closer to the new technological
frontiers, involving Artificial Intelligence (AI), Machine Learning
(ML) and Deep Learning (DL) (Figure 1).
“Images are more than pictures, they are data.”1 The
mantra of modern radiology portrays the potential of this new
understanding of imaging in the new age of precision medicine,
going far beyond diagnosis and having a decisive role in clinical
decision making. In this new and complex context, Cardiology
has been a broad and fertile ground for AI approaches, as many
heterogeneous and sufficiently prevailing diseases (ideal for large
databases), such as heart failure and coronary artery disease, are
yet to be sub-phenotyped in the constant pursuit of increasingly
customized treatments. Besides, problems with acquisition time,
high costs, efficiency and misdiagnosis are commonly observed
and thus expected to be mitigated with the promising new
applications of AI in cardiovascular propaedeutics.2
The first applications of AI in cardiac imaging were
based on the automated quantitative measurement of
anatomical parameters (such as stenosis and vascular
dilatations) and functional parameters (such as ventricular
ejection fraction) that were previously performed
manually and were often considered laborious and
time-consuming. The latest rapidly-evolving applications
include the prediction of myocardial ischemia from
automated coronary Fractional Flow Reserve (FFR) analysis
using computed tomography (CT)3 and identification of
vulnerable plaques by CT angiography using radiomics, a
tool of quantitative evaluation of images based on textural
analysis, that is, heterogeneity of an area of interest from
the distribution on pixels or gray levels of each voxel. Other
recent applications include the development of automatic
reconstruction algorithms, analysis of image quality with DL5
and identification of incidental cardiovascular findings in
CT and Magnetic Resonance Imaging (MRI) examinations.
In addition to these axial methods, potential uses of AI
tools in echocardiography are also comprehensive and
include automated functional assessment (including
ejection fraction and longitudinal strain), quantification
of segmental contractility anomalies, and recognition of
axes and structures with DL techniques.6 Our group has
been a pioneer in research on the subject in the country,
developing studies that include textural analysis of coronary
plaques and aortic valve by radiomics (Figure 2), prognostic
markers in cardiomyopathies with ML7 techniques and
ventricular function evaluation with DL8 tools.
Despite the encouraging results and the growing number
of publications on the subject, there is still a long way to go
before scientific evidence involving AI in cardiac imaging
is implemented in clinical practice. The methodology of
many of the latest studies differs significantly, and some
of them have used mathematical corrections, which can
lead to overly optimistic results — many without external
validation. Another important point to be considered,
when use AI to solve medical image problems, is the
limited number of annotated data available for training.
In most of the cases, this requires the involvement of well-
trained physicians in a time consuming task, which limits
the number of annotated data available. Other limitations
include frequent variations of acquisition protocols (which
may reduce data robustness), as well as the diversity of
the methodologies used to extract features (quantitative
information derived from images) and interpretation of
the statistical models used. Consequently, more studies
are needed to validate the potential of these techniques,
preferably with larger samples, interinstitutional cooperation
and consistent methodology and validation.
Notwithstanding the great enthusiasm of the scientific
community, numerous doubts about the effects of all
this great potential has generated excessive anxiety
among professionals dealing with medical imaging. It is
paramount to state that the use of AI tools in medical
diagnosis is neither a threat nor a strategy to replace the
role of the physician in the cardiovascular propaedeutics.
By providing new diagnostic, predictive and prognostic
data, with potential impact on the individualized therapy
of these patients, such tools undoubtedly represent a
potential strategy to increase the importance and precision
of our work. We believe that AI has great chances of
promoting optimization of workflow and support in the

155
Editorial
Arq Bras Cardiol: Imagem cardiovasc. 2019;32(3):154-156
Araujo-Filho et al.
Artificial Intelligence and Cardiac Imaging
diagnosis in Radiology, as well as reducing the amount of
stress and exhaustion among professionals and improving
the quality of patient care.
Finally, we are sure that the incorporation of AI into cardiac
imaging is not something to be feared or avoided, but discussed,
understood and encouraged, both in multidisciplinary medical
meetings and in the academic training of new professionals.
We believe that all clinical and pathophysiological knowledge
gained throughout our training and career will remain vital for
the balanced use and interpretation of the new data generated
by these tools.9 It is up to us to take a leading role in deciding
where and how to apply all this knowledge.
Figure 1 – Diagram illustrating the basic concepts of articial intelligence, machine learning and deep learning.
Figure 2 – Workow for characterization of coronary plaque using radiomics. First-order features were extracted from a heterogeneous plaque in the anterior
descending artery.
Image acquisition and selection
of the volume of interest
Lesion
segmentation
Histogram
construction and
analysis
Data extraction and
analysis
Relative frequncy
Mitigation
(HU)
Mean
SD
Skewness
Kurtosis
275
190
0.75
1.93
Data prediction
and validation
ARTIFICIAL INTELLIGENCE
Any technique or system that allows
computers to mimic human behavior (feeling,
thinking, acting and adapting)
MACHINE LEARNING
A technique used to provide articial
intelligence with the capacity to learn
DEEP LEARNING
Class of machine learning
algorithms characterized by
the use of complex neural
networks

156
Editorial
Arq Bras Cardiol: Imagem cardiovasc. 2019;32(3):154-156
Araujo-Filho et al.
Artificial Intelligence and Cardiac Imaging
Referências
1. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures,
They Are Data. Radiology. 2016;278(2):563-77.
2. Dey D, Slomka PJ, Leeson P, Comaniciu D, Shrestha S, Sengupta PP, et al.
Artificial Intelligence in Cardiovascular Imaging: JACC State-of-the-Art
Review. J Am Coll Cardiol.2019;73(11):1317-35.
3.
Gaur S, Ovrehus KA, Dey D, Leipsic J, Botker HE, Jensen JM, et al. Coronary plaque
quantification and fractional flow reserve by coronary computed tomography
angiography identify ischaemia-causing lesions. Eur Heart J. 2016;37(15):1220-7.
4. Kolossvary M, Karady J, Szilveszter B, Kitslaar P, Hoffmann U, Merkely
B, et al. Radiomic Features Are Superior to Conventional Quantitative
Computed Tomographic Metrics to Identify Coronary Plaques With
Napkin-Ring Sign. Circ Cardiovasc Imaging.2017;10(12). pii: e006843.
5. Soffer S, Ben-Cohen A, Shimon O, Amitai MM, Greenspan H, Klang E.
Convolutional Neural Networks for Radiologic Images: A Radiologist’s
Guide. Radiology. 2019;290(3):590-606.
6. Alsharqi M, Woodward WJ, Mumith JA, Markham DC, Upton R,
Leeson P. Artificial intelligence and echocardiography. Echo Res Pract.
2018;5(4):R115-R25.
7. Rocon C, Tabassian M, Tavares De Melo MD, Araujo Filho JA, Parga Filho
JR, Hajjar LA, et al. Biventricular imaging markers to predict outcome in
non-compaction cardiomyopathy: a machine learning study. Eur Heart J.
2018;39(suppl_1), ehy566.P6485.
8.
Moreno RA, Rebelo MFSdS, Carvalho T, Assunção AN, Dantas RN,
Val Rd, et al. A combined deep-learning approach to fully automatic
left ventricle segmentation in cardiac magnetic resonance imaging.
Proc. SPIE 10953, Medical Imaging 2019: Biomedical Applications
in Molecular, Structural, and Functional Imaging, 109531Y (15
March 2019).
9. Johnson KW, Torres Soto J, Glicksberg BS, Shameer K, Miotto R,
Ali M, et al. Artificial Intelligence in Cardiology. J Am Coll Cardiol.
2018;71(23):2668-79.