ArticlePDF Available

Automated analysis and detection of abnormalities in transaxial anatomical cardiovascular magnetic resonance images: a proof of concept study with potential to optimize image acquisition

Authors:

Abstract and Figures

The large number of available MRI sequences means patients cannot realistically undergo them all, so the range of sequences to be acquired during a scan are protocolled based on clinical details. Adapting this to unexpected findings identified early on in the scan requires experience and vigilance. We investigated whether deep learning of the images acquired in the first few minutes of a scan could provide an automated early alert of abnormal features. Anatomy sequences from 375 CMR scans were used as a training set. From these, we annotated 1500 individual slices and used these to train a convolutional neural network to perform automatic segmentation of the cardiac chambers, great vessels and any pleural effusions. 200 scans were used as a testing set. The system then assembled a 3D model of the thorax from which it made clinical measurements to identify important abnormalities. The system was successful in segmenting the anatomy slices (Dice 0.910) and identified multiple features which may guide further image acquisition. Diagnostic accuracy was 90.5% and 85.5% for left and right ventricular dilatation, 85% for left ventricular hypertrophy and 94.4% for ascending aorta dilatation. The area under ROC curve for diagnosing pleural effusions was 0.91. We present proof-of-concept that a neural network can segment and derive accurate clinical measurements from a 3D model of the thorax made from transaxial anatomy images acquired in the first few minutes of a scan. This early information could lead to dynamic adaptive scanning protocols, and by focusing scanner time appropriately and prioritizing cases for supervision and early reporting, improve patient experience and efficiency.
This content is subject to copyright. Terms and conditions apply.
Vol.:(0123456789)
1 3
The International Journal of Cardiovascular Imaging (2021) 37:1033–1042
https://doi.org/10.1007/s10554-020-02050-w
ORIGINAL PAPER
Automated analysis anddetection ofabnormalities intransaxial
anatomical cardiovascular magnetic resonance images: aproof
ofconcept study withpotential tooptimize image acquisition
JamesP.Howard1 · SameerZaman1· AarabyRagavan1· KerryHall1· GregLeonard1· SharonSutanto1·
VijayRamadoss1· YousufRazvi1· NickF.Linton1· AnilBharath1· MatthewShun‑Shin1· DanielRueckert1·
DarrelFrancis1· GrahamCole1
Received: 16 August 2020 / Accepted: 29 September 2020 / Published online: 29 October 2020
© The Author(s) 2020
Abstract
The large number of available MRI sequences means patients cannot realistically undergo them all, so the range of sequences
to be acquired during a scan are protocolled based on clinical details. Adapting this to unexpected findings identified early
on in the scan requires experience and vigilance. We investigated whether deep learning of the images acquired in the first
few minutes of a scan could provide an automated early alert of abnormal features. Anatomy sequences from 375 CMR scans
were used as a training set. From these, we annotated 1500 individual slices and used these to train a convolutional neural
network to perform automatic segmentation of the cardiac chambers, great vessels and any pleural effusions. 200 scans were
used as a testing set. The system then assembled a 3D model of the thorax from which it made clinical measurements to
identify important abnormalities. The system was successful in segmenting the anatomy slices (Dice 0.910) and identified
multiple features which may guide further image acquisition. Diagnostic accuracy was 90.5% and 85.5% for left and right
ventricular dilatation, 85% for left ventricular hypertrophy and 94.4% for ascending aorta dilatation. The area under ROC
curve for diagnosing pleural effusions was 0.91. We present proof-of-concept that a neural network can segment and derive
accurate clinical measurements from a 3D model of the thorax made from transaxial anatomy images acquired in the first
few minutes of a scan. This early information could lead to dynamic adaptive scanning protocols, and by focusing scanner
time appropriately and prioritizing cases for supervision and early reporting, improve patient experience and efficiency.
Keywords Artificial intelligence· Machine learning· Cardiac magnetic resonance imaging· Neural networks
Abbreviations
3D Three-dimensional
AUC Area under curve
CMR Cardiovascular magnetic resonance
DCM Dilated cardiomyopathy
HCM Hypertrophic cardiomyopathy
LV Left ventricle
LVEF Left ventricular ejection fraction
LVEDVi LV end-diastolic volume
LVMi Left ventricular mass
ROC Receiver operating characteristic
RVEDVi Right ventricular end-diastolic volume
Background
CMR offers limitless scan planes and a large range of dif-
ferent sequences to characterize the heart in different ways,
but this means that the acquisition process must be selec-
tive because not every patient can have every imaging plane
scanned with every sequence. Efficient clinical practice
therefore operates through standardized “protocols”, which
list the sequences that are most likely to generate images
James P. Howard and Sameer Zaman have contributed equally
towards this work.
Electronic supplementary material The online version of this
article (https ://doi.org/10.1007/s1055 4-020-02050 -w) contains
supplementary material, which is available to authorized users.
* James P. Howard
jphoward@doctors.org.uk
1 Department ofComputing, National Heart andLung
Institute, Imperial College London, Imperial College
Healthcare NHS Trust, London, UK
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
1034 The International Journal of Cardiovascular Imaging (2021) 37:1033–1042
1 3
which confirm or refute important diagnoses for particular
clinical scenarios [1].
However, sometimes the earliest images acquired dur-
ing the scan itself reveal unexpected findings which may
(a) indicate that some previously protocolled images are no
longer required or (b) indicate that further images should
be acquired to confirm or refute a previously unsuspected
diagnosis. Some sequences are best interpreted before gado-
linium contrast has been given, meaning that the opportunity
to use them may be lost if the requirement for them is not
spotted early.
In recent years, deep learning using neural networks has
shown increasing performance in the classification and seg-
mentation of medical imaging data [2]. Recent work has
shown similar precision to human techniques at tasks such as
left ventricular (LV) segmentation [35]. These approaches,
however, have focused on the segmentation of high-quality
LV-dedicated sequences with ideal scan plane orientation.
Within the first couple of minutes of a CMR scan, a
series of transaxial images are routinely acquired, com-
monly termed the “anatomy” sequences. In this study we
investigate whether deep learning methods could analyze
these early images to extract information that could (a) help
a radiographer recognize the need to modify the ongoing
protocol, (b) help identify cases that should be prioritized
for medical supervision or early reporting.
Methods
Data extraction
We extracted 575 sequential CMR scans from 2 manufactur-
ers (Philips Healthcare, Amsterdam, Netherland & Siemens,
Erlangen, Germany) performed across 2 centers. Inclusion
criteria were adult scans in which a bright-blood transaxial
anatomical sequence had been performed and a valid final
report which included the following BSA-indexed (suffix
i) measurements as a minimum: LV end-diastolic volume
(LVEDVi), LV mass (LVMi) and RV end-diastolic volume
(RVEDVi). We did not exclude scans with artefact, e.g. due
to atrial fibrillation or breath-holding difficulties. We also
extracted the following measures from the report where
present: the diameter of the ascending aorta; the presence
of hypertrophic cardiomyopathy (HCM), the presence of
dilated cardiomyopathy (DCM) and the presence of pleural
effusions.
The CMR scans were then randomly assigned to differ-
ent datasets, each serving a specific purpose (Fig.1; Study
Design). 200 scans were assigned to the “testing dataset”.
This dataset was used to report the final accuracy of the
system and was not shown to the neural network at any stage
of its training. The remaining 375 scans were assigned to
the “training & validation dataset”. This dataset was used
to train the neural network. 75% of these scans were used to
directly train the neural network (the “training dataset”), and
25% were used to continually appraise the performance of
the network during development (the “validation dataset”).
Ethical approval was gained from the Health Regulatory
Agency (Integrated Research Application System identifier
243,023).
Data processing andlabelling
The axial anatomical images were isolated from each scan.
The acquisition parameters for these sequences are shown in
Appendix1. 1500 slices were then chosen at random, across
the training and validation sets and were then labelled by
clinicians using custom-designed software to draw around
the following anatomical features, if present: ascending
aorta, aortic arch, descending aorta, left atrium, left ventri-
cle, pericardial effusion, pleural effusion, right atrium and
right ventricle. Each slice was then resampled to 560 by 560
pixels, with zero-padding for non-square acquired images.
Neural network design andtraining
The neural network design chosen was an adapted version of
the HRNet architecture [6]. Modifications were made so the
network could receive single-channel (grayscale) images and
output 9 feature maps, corresponding to the possible identi-
ties of each pixel (aorta, left atrium, left ventricular wall,
left ventricular cavity, pulmonary artery, pleural effusion,
right atrium, right ventricular cavity) or a final ‘other’ (back-
ground) class. The network was trained until the validation
loss plateaued (23 epochs), and the training process was
augmented with random rescaling, rotation, shearing and
translation. Loss was calculated over batches of 20 images
by using the categorical cross-entropy loss function, and
weights were updated using the Ranger optimizer (a com-
bination of RAdam [7] and Lookahead [8]) with a learning
rate of 0.001. The optimizer, choice of data augmentations
and learning rate were tuned with reference to the valida-
tion loss. Programming was performed using the Python
programming language with the Pytorch framework [9].
Training was performed on 2 GeForce RTX Titan graphical
processing units (NVIDIA, Santa Clara, California).
3D heart model reconstruction
The segmentation predictions of the neural network were
converted into predictions of parameters such as cardiac
chamber sizes by analyzing each slice within the axial ana-
tomical planning sequence and then re-assembling them into
a 3D heart model (Fig.2). Specifically, each slice was fed
into the neural network to yield a prediction of the identity
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
1035The International Journal of Cardiovascular Imaging (2021) 37:1033–1042
1 3
of each pixel (left ventricle, pleural effusion, etc.). The
slices were then resized according to the viewport informa-
tion embedded in the DICOM file and the thickness of the
slice, before combining the stacked series of slices into a
3D model through trilinear interpolation. This model was
then analyzed by calculating the volume of each structure
present to give a value in ml. These were then scaled for
body surface area to yield a value in ml/cm3. Finally, these
volumes were adjusted to the “final” analyzed volumes by
fitting a linear regression model separately for each scan-
ner manufacturer: this allowed a bias to be introduced to
account for the different acquisition parameters between
the two manufacturers. Models were fitted for estimating
LVEDVi, RVEDVi, LVMi, BSA-indexed LV mass-to-vol-
ume ratio, and BSA-indexed ascending aortic diameter. This
allowed compensation for any systematic error associated
with the use of differences in sequences employed by dif-
ferent manufacturers.
For each measure, published BSA-indexed cut-offs
were used to classify abnormal cases: LVEDVi > 94ml/m2
for LV dilatation, RVEDVi > 98ml/m2 for RV dilatation,
LVMi > 83.5g/ml for LV hypertrophy [10], and > 0.84 for
an abnormal LV mass-to-volume ratio [11].
Endpoints andstatistical analysis
The segmentation performance of the neural network was
assessed using the Dice coefficient, defined as:
where for each class, TP refers to true positives (pixels clas-
sified by both the network and expert as of that class), FP as
false positives (pixels classified by the network as that class,
incorrectly) and FN as false negatives (pixels classified by
the network as not that class, incorrectly).
The diagnostic performance of the network was reported
for continuous outcomes using Pearson’s r, along with the
accuracy, sensitivity and specificity. Cohen’s Kappa (κ) was
also reported as a measure of accuracy resistant to class
Dice
=
2
TP
2TP +FP +FN
Source MRI studies
575 studies across 2 MRI manufacturers
Test set
200 studies
Training & Validation set
375 studies
Training
Evaluation of diagnostic performance:
LV & RV volumes
LV mass & mass-to-volume ratio
HCM & DCM pick-up
Pleural effusions
Ascending aorta dimensions
SEGMENTATION
NETWORK
DEVELOPMENT
DATA
COLLECTION
DIAGNOSTIC
ASSESSMENT
Evaluation of
segmentation performance
Dice coefficient
Axial anatomical
planning
sequences
Cardiac
chambers
Great vessels
& pleural effusions
SEGMENTATION NETWORK
SOURCE DATA
Model training
Training set
289 studies
1156 slices
Validation set
86 studies
344 slices
3D THORAX RECONSTRUCTION
WOLFATADNGISEDYDUTS
Fig. 1 Study design flow chart and data flow. The aim of this study
was to use artificial intelligence (AI) to create a 3D model of the
thorax for each patient using the anatomy images acquired within
the first few minutes of a cardiac MRI scan. This could be used to
make measurements and provide useful diagnostic information within
minutes of a scan starting. The study comprised 575 studies across 2
manufacturers, which were split into a training and testing set. The
training set was used to train an AI which could segment each slice
of the anatomy sequence. Then, each scan in the testing set was ana-
lyzed to allow creation of a 3D model of a patient’s thorax. This 3D
model was then analyzed. 3D = 3-dimensional. DCM = dilated cardio-
myopathy. HCM = hypertrophic cardiomyopathy. LV = left ventricle,
RV = right ventricle
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
1036 The International Journal of Cardiovascular Imaging (2021) 37:1033–1042
1 3
imbalance. For categorical outcomes, we reported the area
under the curve (AUC) for the receiver operating character-
istic (ROC), along with the accuracy and Cohen’s Kappa for
the optimal cut-off. Statistical analysis was performed using
R software (R Foundation, Vienna, Austria) [12].
Results
Dataset
The baseline characteristics across the training & validation
dataset and the testing dataset are shown in Table1.
1500 slices underwent segmentation labelling, 1156 of
which were within the training set and 344 within the testing
set. Table2 outlines the contents of the slices.
Segmentation performance
The mean Dice coefficient across all categories for the test-
ing set was 0.910. The category-wise Dice coefficients are
shown in Table2 and ranged between 0.809 for the LV cav-
ity to 0.929 for the aorta.
Examples of the human-supplied labels and the predic-
tions of the neural network are shown in Fig.3. In contrast,
the slice where the neural network made the greatest seg-
mentation error is shown in the Appendix.
Diagnostic model performance
By combining every slice from a scan’s axial anatomic plan-
ning sequence, we were able to construct a 3D model of a
patient’s thorax (Fig.2) from which calculations could be
S
S
T
T
E
E
P
P
2
2
Segment every slice
S
S
T
T
E
E
P
P
3
3
Combine into 3D model
Anterior view
Posterior view
S
S
T
T
E
E
P
P
1
1
Acquire low-resolution axial stack
N
N
e
e
u
u
r
r
a
a
l
l
n
n
e
e
t
t
w
w
o
o
r
r
k
k
I
I
n
n
t
t
e
e
r
r
p
p
o
o
l
l
a
a
t
t
i
i
o
o
n
n
Fig. 2 Reconstruction of a 3D heart is possible by segmenting each
slice within the axial anatomy stack. The slices can be combined
through trilinear interpolation into a 3D model of the cardiac cham-
bers and great arteries. This model can be used to make anatomical
measurements, such as end-diastolic volumes and vessel diameters
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
1037The International Journal of Cardiovascular Imaging (2021) 37:1033–1042
1 3
made of chamber size but also for presence of abnormal
features, such as pleural effusions.
BSA-indexed left ventricular end diastolic volume
(LVEDVi) predicted by the network on the testing set cor-
related with the measures from the final report (R2 = 0.76,
p < 0.0001, Fig.4). The network was 90.5% accurate in iden-
tifying LV dilatation (κ = 0.75), with a corresponding sensi-
tivity of 84.3% and specificity of 92.6%. Within the testing
dataset, 10 studies were from patients with dilated cardio-
myopathy, and the network correctly identified 9 (90%) of
these as abnormal (95% CI 59.6% to 98.2%).
BSA-indexed right ventricular end diastolic volume
(RVEDVi) predicted by the network on the testing set cor-
related with the measures from the final report (R2 = 0.53,
p < 0.0001, Fig. 4). The network was 85.5% accurate in
identifying RV dilatation (κ = 0.62), with a corresponding
sensitivity of 80% and specificity of 87.1%.
BSA-indexed left ventricular mass (LVMi) predicted by
the network on the testing set correlated with the measures
from the final report (R2 = 0.74, p < 0.0001, Fig.5). The
network was 85% accurate in identifying LV hypertrophy
= 0.64), with a corresponding sensitivity of 78.9% and
specificity of 87.4%. LV mass:volume ratio predicted by the
network on the testing set correlated with the measures from
the final report (R2 = 0.5, p < 0.0001, Fig. 5). The network
was 71% accurate in identifying a raised LV mass:volume
ratio (κ = 0.41), with a corresponding sensitivity of 68.9%
and specificity of 74.4%. Within the testing dataset, 10 stud-
ies were from patients with hypertrophic cardiomyopathy,
and the network correctly identified all (100%) of these as
abnormal (95% CI 72.3% to 100%).
Only 54 of the 200 testing dataset cases provided ascend-
ing aorta diameters. BSA-indexed ascending aorta diameters
predicted by the network on the testing set correlated with
the measures from the final report (R2 = 0.82, p < 0.0001,
Fig.6). The network was 94.4% accurate in identifying
ascending aortic dilatation (κ = 0.79), with a corresponding
sensitivity of 100% and specificity of 93.6%.
The ROC curve AUC for diagnosing pleural effusions by
quantifying pleural fluid was 0.906 (see online Appendix).
Discussion
This study is a proof-of-concept, that accurate important
diagnostic information can be derived from axial anatomy
images obtained at the start of a CMR scan. These results are
consistent across two different scanner manufacturers. This
Table 1 Characteristics of the included studies in the training and
testing sets. Values are mean (standard deviation) for continuous vari-
ables, and count (percentage) for categorical variables
LVEF left ventricular ejection fraction, LVEDVi Body-surface area-
indexed left ventricular end-diastolic volumes, RVEDVi Body-surface
area-indexed right ventricular end-diastolic volumes, LVMi Body-
surface area-indexed right ventricular end-diastolic volumes, HCM
hypertrophic cardiomyopathy, DCM dilated cardiomyopathy
Training and valida-
tion set
Testing set
Number of studies/patients 375 200
Manufacturer
Siemens 221 125
Philips 154 75
Male 229 (61.1) 115 (57.5)
Age (years) 55.7 (17.8) 56.7 (17.5)
LVEF (%) 55.6 (17.5) 55.7 (17.8)
LVEDVi (ml/m2) 81.4 (34.6) 81.0 (38.1)
RVEDVi (ml/m2) 80.1 (25.9) 78.8 (26.6)
LVM i (g/m2) 77.6 (30.8) 76.4 (33.2)
Ascending aorta (mm/m2) 17.1 (4.2) 16.8 (4.1)
HCM 25 (6.7) 10 (5.0)
DCM 22 (5.9) 10 (5.0)
Pleural effusions 88 (23.5) 42 (21.0)
Table 2 Slices included in the
training and testing sets
Values are n (%). The Dice coefficient reflects the accuracy of the neural network on the validation set
Feature True presence within a slice Network perfor-
mance (Dice coef-
ficient)
Training set (n = 1156) Validation set (n = 344)
Aorta 1059 (91.6) 316 (91.9) 0.929
Left atrium 404 (34.9) 116 (33.7) 0.925
LV cavity 558 (48.3) 151 (44.1) 0.809
LV wall 587 (50.8) 167 (48.6) 0.884
Pulmonary artery 182 (15.8) 65 (18.9) 0.907
Pleural effusion 347 (30.9) 128 (37.2) 0.890
Right atrium 481 (41.6) 139 (40.4) 0.924
RV cavity 616 (53.3) 178 (51.7) 0.910
Background 1156 (100.0) 344 (100.0) 0.995
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
1038 The International Journal of Cardiovascular Imaging (2021) 37:1033–1042
1 3
system could allow technicians performing the scan to be
signposted to unexpected pathology, help direct optimized
image acquisition for the remainder of the scan, prioritize
supervision of scans by reporters, and prioritize scans for
urgent reporting.
Earlier & automated diagnosis – a comparison with previ-
ous approaches.
Neural networks are now rivalling and surpassing
humans for cardiac chamber segmentation and quantifica-
tion [3, 4]; a situation in which images are acquired in a
dedicated conventional manner for all patients. However,
the aim of our study was different in two ways.
First, the network in this study has the potential to
identify extra-cardiac diagnoses such as aortic dilatation
and pleural effusions. Being able to identify findings may
allow an adaptive approach to scanning protocols which
avoids recalling patients for additional images and even
gives technicians performing the scan reassurance that
additional images are not required.
Fig. 3 A Human label and the
AI’s segmentation correspond-
ing prediction. The figure
showed a comparison between
the human labels of a particu-
lar axial anatomical planning
slice, and those predicted by the
network for that slice
Human opinionComputer predicon
Axial anatomy slice
Neural
Network
Expert
Labelling
Difference appraised with
Dice coefficient
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
1039The International Journal of Cardiovascular Imaging (2021) 37:1033–1042
1 3
Second, previous studies have aimed to improve absolute
quantification of chamber size and function and have there-
fore been trained to process the high-quality cine imaging
which reporting clinicians currently use. Such sequences
differ from the anatomical stack generated and processed in
this paper in several important ways.
Accuracy of automated ventricular volume measurements derived from from axial anatomy images
R2:0.76
Accuracy: 90.5%
Sensitivity: 84.3%
Specificity: 92.6%
R
2
:
0.7
6
cc
r
a
cy: 90.5
%
S
ensitivit
y
: 84.3%
S
pecificity: 92.6%
10
30
100
300
10 30 100300
Clinically reported value (ml / m2 )
Axial anatomy prediction (ml / m2 )
DCM
No
Yes
Manufacturer
Philips
Siemens
LVEDVi
R2:0.53
Accuracy: 85.5%
Sensitivity: 80%
Specificity: 87.1%
R
2
:
0
.5
3
A
ccu
r
acy: 85.5
%
S
ensitivit
y
: 80
%
S
pecificity: 87.1
%
10
30
100
300
10 30 100 300
Clinically reported value (ml / m2 )
Axial anatomy prediction (ml / m2 )
Manufacturer
Philips
Siemens
RVEDVi
Fig. 4 The relationship between the neural network’s chamber size
predictions versus those extracted from the final report. Predictions
are shown for BSA-indexed left ventricular end-diastolic volume
(LVEDVi) and BSA-indexed right ventricular end-diastolic volume
(RVEDVi). Green zones and red zones symbolize areas of the plot
where the neural network agrees and disagrees with the values from
the full scan, respectively. Patients with known dilated cardiomyopa-
thy (DCM) are shown as crosses on the LVEDVi plot
Accuracy of automated left ventricular mass measurements derived from axial anatomy images
R2:0.74
Accuracy: 85%
Sensitivity: 78.9%
Specificity: 87.4%
R
2
:
0
.7
4
Accu
r
a
c
y
: 85%
S
ensitivit
y
: 78.9
%
S
pecificit
y
: 87.4
%
10
30
100
300
10 30 100300
Clinically reported value (g / m2 )
Axial anatomy prediction (g / m2 )
Philips
Siemens
LVMi
R2:0.5
Accuracy: 71%
Sensitivity: 68.9%
Specificity: 74.4%
R
2
:
0
.5
A
cc
u
r
ac
y
: 71
%
S
ensitivit
y
: 68.9
%
S
pecificity: 74.4
%
0.3
1.0
3.0
0.31.0 3.0
Clinically reported value (g / ml)
Axial anatomy prediction (g / ml)
HCM
No
Yes
Manufacturer
Philips
Siemens
LV mass:volume ratio
Fig. 5 The relationship between the neural network’s left ventricular
mass predictions versus those extracted from the final report. Predic-
tions are shown for BSA-indexed left ventricular mass (LVMi) and
the left ventricular mass:volume ratio, where volume is derived from
the neural network’s prediction of the left ventricular end diastolic
volume. Green zones and red zones symbolize areas of the plot where
the neural network agrees and disagrees with the values from the full
scan, respectively. Patients with known hypertrophic cardiomyopathy
(HCM) are shown as crosses on the LVMi plot
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
1040 The International Journal of Cardiovascular Imaging (2021) 37:1033–1042
1 3
First, the dedicated sequences’ scan planes are ideally ori-
entated with respect to the patient’s heart to ensure radial func-
tion is within rather than through the plane of imaging and
minimizes partial voluming. Second, the dedicated sequences
are acquired with less spacing, allowing more voxels per ven-
tricle. Third, the dedicated sequences take considerably longer
to acquire: they are acquired over 8–15s per slice, of which
there may be up to 10 slices per sequence [3].
In contrast, the entire axial anatomical sequences can be
acquired in 3–4 breath holds. However, they are of relatively
large slice thickness, and are orientated axially with no refer-
ence to the patient’s heart. Given this, the ability to acquire
accurate diagnostic information early on in the scan has
the potential to triage the application of further sequences
during the scan. For example, patients found to have LVH
and pleural effusions on axial anatomy images may have a
diagnosis of cardiac amyloid in which pre-contrast T1 maps
would be useful [13]. Dedicated aortic imaging may be too
laborious for routine use on each patient but could be reli-
ably targeted to those who need it by our work.
Implications forreporting prioritization,
supervision andpatient safety
A system providing automated diagnosis during the earli-
est stages of a cardiac MRI scan would not only be useful
for ensuring scans are correctly protocolled but would also
allow physicians to prioritize the supervision and reporting
of those scans most likely to be abnormal.
Patients with pleural effusions, for example, may have
limited tolerance for lying flat in the scanner. Their identi-
fication at the earliest stages of could increase vigilance of
these more vulnerable patients, and alteration of sequences
to allow smaller breath holds, motion-corrected free-breath-
ing sequences and accelerated protocols to minimize scan
time.
Scans shown to contain unexpected pathologies may also
be flagged for expedited review and reporting by physicians.
For example, an outpatient screening CMR scan in a low
risk patient might be identified as demonstrating unexpected
marked LV dilatation with pleural effusions. Such patients
at risk of decompensation and could be identified for early
reporting and follow-up.
Study limitations
The neural network described in this study is not 100% accu-
rate. Even the most accurate measurement (ascending aortic
diameter) is only 94% accurate on our dataset. However, the
correlation between the neural network’s predictions and the
true measurements for the measures examined ensure that
the extreme biological values associated with disease are
Fig. 6 The relationship between
the neural network’s BSA-
indexed ascending aortic diam-
eter predictions versus those
extracted from the final report.
Green zones and red zones sym-
bolize areas of the plot where
the neural network agrees and
disagrees with the values from
the full scan, respectively
Accuracy of automated ascending aorta diameter measurement
derived from axial anatomy images
R2:0.82
Accuracy: 94.4%
Sensitivity: 100%
Specificity: 93.6%
R
2
:
0
.
82
Accu
r
a
c
y
: 94.4
%
S
ensitivit
y
: 100
%
S
peci
f
icit
y
: 93.6
%
0
10
20
30
40
010203
04
0
Clinically reported value (mm /
m
2
)
Axial anatomy prediction (mm / m2 )
Manufacturer
Philips
Siemens
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
1041The International Journal of Cardiovascular Imaging (2021) 37:1033–1042
1 3
more consistently correctly identified by the network (100%
of hypertrophic cardiomyopathy cases, 90% of dilated car-
diomyopathy cases).
It is difficult to ascertain to what extent the errors in the
neural network’s predictions are due to inaccurate segmenta-
tion by the network, versus limitations inherent to estimating
volumes using anatomy sequence slices. The latter could be
estimated by calculating volumes using expert labels across
the testing dataset, although this would require every myo-
cardial slice in these data to be labelled. Whilst this dataset
would be many times larger than the current dataset used to
train the network, we hope to address this question in the
future.
Furthermore, previous studies have shown the coefficient
of variation is over 10% for estimating left ventricular vol-
umes, even when assessed by the same doctor in the same
patient using dedicated left ventricular sequences [3]. In this
study, we have compared the performance of this network
against human observers behaving clinically, and therefore
this 10% variation inherently sets an upper ceiling on the
correlation coefficient obtainable by the network—it cannot
correlate with the human observers better than the human
observers correlate with themselves.
With all deep learning studies, there is concern that the
findings in this study may not generalize to a wider popula-
tion [14]. This can be due to a phenomenon of “overfitting”,
where the neural network is highly accurate at processing
images on which it was trained but performs much less well
on unseen “real world” examples. To try and mitigate these
concerns, we took several approaches. First, the performance
is reported on a test set which was only assessed after train-
ing the neural network. Second, the dataset we assembled
was from two different hospitals across multiple reporting
physicians. Third, the dataset comprises scans across both
Siemens and Philips scanners, with the correlation plots
showing similar accuracies for both manufacturers. Finally,
we are releasing the neural network with this manuscript for
use online, so that its performance can be assessed by any
interested academic or clinician.
Conclusion
This proof-of-concept study demonstrated that a neural net-
work can accurately reconstruct a 3-dimensional model of
the heart and major vessels from transaxial anatomy images
acquired in the first few minutes of a CMR study. Our system
is able to accurately quantify cardiac chamber size, aortic
diameter and presence of pleural effusions. We have made
trained neural networks publicly available for use.
Acknowledgements This study was supported by the NIHR Imperial
Biomedical Research Centre (BRC). The views expressed are those of
the author(s) and not necessarily those of the NIHR or the Department
of Health and Social Care.
Author contributions JPH, DF and GC conceived and designed the
study. JH, SZ, AR, KH, GL, SS, VR, YR extracted and processed the
data. NL, AB, MS, DR, DF and GC drafted and revised the manuscript.
Funding JPH is funded by the Wellcome Trust (212183/Z/18/Z).
Data availability The trained neural network and inference code will be
made available on the author’s website at https ://james .dev—the source
data used to train these networks, however, are not publicly available
due to ethical restrictions.
Compliance with ethical standards
Conflict of interest The authors declare that there are no competing
interests.
Ethical approval Ethical approval was gained from the Health Regu-
latory Agency (Integrated Research Application System identifier
243023).
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. Kramer CM, Barkhausen J, Flamm SD, Kim RJ, Nagel E (2013)
Standardized cardiovascular magnetic resonance (CMR) pro-
tocols 2013 update. J Cardiovasc Magn Reson. https ://doi.
org/10.1186/1532-429X-15-91
2. Howard JP, Fisher L, Shun-Shin MJ, Keene D, Arnold AD,
Ahmad Y etal (2019) Cardiac rhythm device identification using
neural networks. JACC Clin Electrophysiol 5:576–586
3. Bhuva AN, Bai W, Lau C, Davies RH, Ye Y, Bulluck H etal
(2019) A multicenter, scan-rescan, human and machine learn-
ing CMR study to test generalizability and precision in imag-
ing biomarker analysis. Circ Cardiovasc Imaging. https ://doi.
org/10.1161/CIRCI MAGIN G.119.00921 4
4. Ruijsink B, Puyol-Antón E, Oksuz I, Sinclair M, Bai W, Schna-
bel JA etal (2019) Fully automated, quality-controlled cardiac
analysis from CMR. JACC Cardiovasc Imaging. https ://doi.
org/10.1016/j.jcmg.2019.05.030
5. Bai W, Sinclair M, Tarroni G, Oktay O, Rajchl M, Vaillant G
etal (2018) Automated cardiovascular magnetic resonance image
analysis with fully convolutional networks. J Cardiovasc Magn
Reson. https ://doi.org/10.1186/s1296 8-018-0471-x
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
1042 The International Journal of Cardiovascular Imaging (2021) 37:1033–1042
1 3
6. Sun K, Zhao Y, Jiang B, Cheng T, Xiao B, Liu D etal (2019)
High-resolution representations for labeling pixels and regions.
https ://arxiv .org/abs/1904.04514 . Accessed 30 Aug 2019
7. Liu L, Jiang H, He P, Chen W, Liu X, Gao J etal (2019) On the
variance of the adaptive learning rate and beyond. https ://arxiv
.org/abs/1908.03265 . Accessed 10 Oct 2019
8. Zhang MR, Lucas J, Hinton G, Ba J (2019) Lookahead optimizer:
k steps forward, 1 step back. https ://arxiv .org/abs/1907.08610 .
Accessed 10 Oct 2019
9. Paszke A, Gross S, Chintala S, Chanan G, Yang E, DeVito Z etal
(2017) Automatic differentiation in PyTorch. 2017. https ://openr
eview .net/forum ?id=BJJsr mfCZ. Accessed 4 Sep 2019
10. Kawel-Boehm N, Maceira A, Valsangiacomo-Buechel ER, Vogel-
Claussen J, Turkbey EB, Williams R etal (2015) Normal values
for cardiovascular magnetic resonance in adults and children. J
Cardiovasc Magn Reson 7:775–782
11. Petersen SE, Aung N, Sanghvi MM, Zemrak F, Fung K, Paiva JM
etal (2017) Reference ranges for cardiac structure and function
using cardiovascular magnetic resonance (CMR) in Caucasians
from the UK Biobank population cohort. J Cardiovasc Magn
Reson 19:18
12. R Core Team (2014) R: a language and environment for statistical
computing [Internet]. 2014. https ://www.r-proje ct.org/
13. Karamitsos TD, Piechnik SK, Banypersad SM, Fontana M, Ntusi
NB, Ferreira VM etal (2013) Noncontrast T1 mapping for the
diagnosis of cardiac amyloidosis. JACC Cardiovasc Imaging
6:488–497
14. Stupple A, Singerman D, Celi LA (2019) The reproducibility cri-
sis in the age of digital medicine. npj Digit Med. Springer Nature
2:2
Publisher’s Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
1.
2.
3.
4.
5.
6.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at
onlineservice@springernature.com
... Aims In this project, we aim to estimate PATV from clinical CMRI. We focus on axial T2-weighted rapid gradient-echo scans which are routinely acquired in the clinic and in which CNNs have been used to accurately segment several cardiac structures [11]. We initially create a non-learned semi-automatic image processing protocol to segment PAT and generate training data for the proposed CNN. ...
... Preprocessing Ground truth PAT segmentation was semi-automatically performed in 70 training images (Fig 1). As the first step, we identified the cardiac chambers and left ventricular (LV) myocardium using HRNet, an existing CNN trained in T2-weighted images with a similar contrast [11]. Using these segmentations, we created a bounding box around the heart. ...
Preprint
Full-text available
Background: Increased pericardial adipose tissue (PAT) is associated with many types of cardiovascular disease (CVD). Although cardiac magnetic resonance images (CMRI) are often acquired in patients with CVD, there are currently no tools to automatically identify and quantify PAT from CMRI. The aim of this study was to create a neural network to segment PAT from T2-weighted CMRI and explore the correlations between PAT volumes (PATV) and CVD outcomes and mortality. Methods: We trained and tested a deep learning model, PAT-CNN, to segment PAT on T2-weighted cardiac MR images. Using the segmentations from PAT-CNN, we automatically calculated PATV on images from 391 patients. We analysed correlations between PATV and CVD diagnosis and 1-year mortality post-imaging. Results: PAT-CNN was able to accurately segment PAT with Dice score/ Hausdorff distances of 0.74 +- 0.03/27.1 +- 10.9~mm, similar to the values obtained when comparing the segmentations of two independent human observers ($0.76 +- 0.06/21.2 +- 10.3~mm$). Regression models showed that, independently of sex and body-mass index, PATV is significantly positively correlated with a diagnosis of CVD and with 1-year all cause mortality (p-value < 0.01). Conclusions: PAT-CNN can segment PAT from T2-weighted CMR images automatically and accurately. Increased PATV as measured automatically from CMRI is significantly associated with the presence of CVD and can independently predict 1-year mortality.
... Natural language processing can automate clinical documentation by text summarization [11] and detect diagnoses from scan reports [12]. Computer vision algorithms can detect lesions from radiological scans, make measurements, and pick-up incidental findings, saving time and reducing error [13,14]. Robots trained by human operators could perform procedures such as venepuncture, and ultrasound [15,16]. ...
Article
Full-text available
Background Artificial intelligence (AI) technologies are increasingly used in clinical practice. Although there is robust evidence that AI innovations can improve patient care, reduce clinicians’ workload and increase efficiency, their impact on medical training and education remains unclear. Methods A survey of trainee doctors’ perceived impact of AI technologies on clinical training and education was conducted at UK NHS postgraduate centers in London between October and December 2020. Impact assessment mirrored domains in training curricula such as ‘clinical judgement’, ‘practical skills’ and ‘research and quality improvement skills’. Significance between Likert-type data was analysed using Fisher’s exact test. Response variations between clinical specialities were analysed using k-modes clustering. Free-text responses were analysed by thematic analysis. Results Two hundred ten doctors responded to the survey (response rate 72%). The majority (58%) perceived an overall positive impact of AI technologies on their training and education. Respondents agreed that AI would reduce clinical workload (62%) and improve research and audit training (68%). Trainees were skeptical that it would improve clinical judgement (46% agree, p = 0.12) and practical skills training (32% agree, p < 0.01). The majority reported insufficient AI training in their current curricula (92%), and supported having more formal AI training (81%). Conclusions Trainee doctors have an overall positive perception of AI technologies’ impact on clinical training. There is optimism that it will improve ‘research and quality improvement’ skills and facilitate ‘curriculum mapping’. There is skepticism that it may reduce educational opportunities to develop ‘clinical judgement’ and ‘practical skills’. Medical educators should be mindful that these domains are protected as AI develops. We recommend that ‘Applied AI’ topics are formalized in curricula and digital technologies leveraged to deliver clinical education.
Chapter
Cardiac magnetic resonance (CMR) is the gold standard for quantification of cardiac volumes, function, and blood flow. Tailored MR pulse sequences define the contrast mechanisms, acquisition geometry and timing which can be applied during CMR to achieve unique tissue characterisation. It is impractical for each patient to have every possible acquisition option. We target the aortic valve in the three-chamber (3-CH) cine CMR view. Two major types of anomalies are possible in the aortic valve. Stenosis: the narrowing of the valve which prevents an adequate outflow of blood, and insufficiency (regurgitation): the inability to stop the back-flow of blood into the left ventricle. We develop and evaluate a deep learning system to accurately classify aortic valve abnormalities to enable further directed imaging for patients who require it. Inspired by low level image processing tasks, we propose a multi-level network that generates heat maps to locate the aortic valve leaflets’ hinge points and aortic stenosis or regurgitation jets. We trained and evaluated all our models on a dataset of clinical CMR studies obtained from three NHS hospitals (n = 1,017 patients). Our results (mean accuracy = 0.93 and F1 score = 0.91), show that an expert-guided deep learning-based feature extraction and a classification model provide a feasible strategy for prescribing further, directed imaging, thus improving the efficiency and utility of CMR scanning.KeywordsCardiac MRIAortic valveExplainabilityMachine learning
Article
Full-text available
Purpose: To assess whether the semisupervised natural language processing (NLP) of text from clinical radiology reports could provide useful automated diagnosis categorization for ground truth labeling to overcome manual labeling bottlenecks in the machine learning pipeline. Materials and methods: In this retrospective study, 1503 text cardiac MRI reports from 2016 to 2019 were manually annotated for five diagnoses by clinicians: normal, dilated cardiomyopathy (DCM), hypertrophic cardiomyopathy, myocardial infarction (MI), and myocarditis. A semisupervised method that uses bidirectional encoder representations from transformers (BERT) pretrained on 1.14 million scientific publications was fine-tuned by using the manually extracted labels, with a report dataset split into groups of 801 for training, 302 for validation, and 400 for testing. The model's performance was compared with two traditional NLP models: a rule-based model and a support vector machine (SVM) model. The models' F1 scores and receiver operating characteristic curves were used to analyze performance. Results: After 15 epochs, the F1 scores on the test set of 400 reports were as follows: normal, 84%; DCM, 79%; hypertrophic cardiomyopathy, 86%; MI, 91%; and myocarditis, 86%. The pooled F1 score and area under the receiver operating curve were 86% and 0.96, respectively. On the same test set, the BERT model had a higher performance than the rule-based model (F1 score, 42%) and SVM model (F1 score, 82%). Diagnosis categories classified by using the BERT model performed the labeling of 1000 MR images in 0.2 second. Conclusion: The developed model used labels extracted from radiology reports to provide automated diagnosis categorization of MR images with a high level of performance.Keywords: Semisupervised Learning, Diagnosis/Classification/Application Domain, Named Entity Recognition, MRI Supplemental material is available for this article. © RSNA, 2021.
Preprint
Full-text available
The aortic vessel tree is composed of the aorta and its branching arteries, and plays a key role in supplying the whole body with blood. Aortic diseases, like aneurysms or dissections, can lead to an aortic rupture, whose treatment with open surgery is highly risky. Therefore, patients commonly undergo drug treatment under constant monitoring, which requires regular inspections of the vessels through imaging. The standard imaging modality for diagnosis and monitoring is computed tomography (CT), which can provide a detailed picture of the aorta and its branching vessels if combined with a contrast agent, resulting in a CT angiography (CTA). Optimally, the whole aortic vessel tree geometry from consecutive CTAs, are overlaid and compared. This allows to not only detect changes in the aorta, but also more peripheral vessel tree changes, caused by the primary pathology or newly developed. When performed manually, this reconstruction requires slice by slice contouring, which could easily take a whole day for a single aortic vessel tree and, hence, is not feasible in clinical practice. Automatic or semi-automatic vessel tree segmentation algorithms, on the other hand, can complete this task in a fraction of the manual execution time and run in parallel to the clinical routine of the clinicians. In this paper, we systematically review computing techniques for the automatic and semi-automatic segmentation of the aortic vessel tree. The review concludes with an in-depth discussion on how close these state-of-the-art approaches are to an application in clinical practice and how active this research field is, taking into account the number of publications, datasets and challenges.
Article
Full-text available
Background: Automated analysis of cardiac structure and function using machine learning (ML) has great potential, but is currently hindered by poor generalizability. Comparison is traditionally against clinicians as a reference, ignoring inherent human inter- and intraobserver error, and ensuring that ML cannot demonstrate superiority. Measuring precision (scan:rescan reproducibility) addresses this. We compared precision of ML and humans using a multicenter, multi-disease, scan:rescan cardiovascular magnetic resonance data set. Methods: One hundred ten patients (5 disease categories, 5 institutions, 2 scanner manufacturers, and 2 field strengths) underwent scan:rescan cardiovascular magnetic resonance (96% within one week). After identification of the most precise human technique, left ventricular chamber volumes, mass, and ejection fraction were measured by an expert, a trained junior clinician, and a fully automated convolutional neural network trained on 599 independent multicenter disease cases. Scan:rescan coefficient of variation and 1000 bootstrapped 95% CIs were calculated and compared using mixed linear effects models. Results: Clinicians can be confident in detecting a 9% change in left ventricular ejection fraction, with greater than half of coefficient of variation attributable to intraobserver variation. Expert, trained junior, and automated scan:rescan precision were similar (for left ventricular ejection fraction, coefficient of variation 6.1 [5.2%-7.1%], P=0.2581; 8.3 [5.6%-10.3%], P=0.3653; 8.8 [6.1%-11.1%], P=0.8620). Automated analysis was 186× faster than humans (0.07 versus 13 minutes). Conclusions: Automated ML analysis is faster with similar precision to the most precise human techniques, even when challenged with real-world scan:rescan data. Assessment of multicenter, multi-vendor, multi-field strength scan:rescan data (available at www.thevolumesresource.com) permits a generalizable assessment of ML precision and may facilitate direct translation of ML to clinical practice.
Article
Full-text available
Objectives: This study sought to develop a fully automated framework for cardiac function analysis from cardiac magnetic resonance (CMR), including comprehensive quality control (QC) algorithms to detect erroneous output. Background: Analysis of cine CMR imaging using deep learning (DL) algorithms could automate ventricular function assessment. However, variable image quality, variability in phenotypes of disease, and unavoidable weaknesses in training of DL algorithms currently prevent their use in clinical practice. Methods: The framework consists of a pre-analysis DL image QC, followed by a DL algorithm for biventricular segmentation in long-axis and short-axis views, myocardial feature-tracking (FT), and a post-analysis QC to detect erroneous results. The study validated the framework in healthy subjects and cardiac patients by comparison against manual analysis (n = 100) and evaluation of the QC steps' ability to detect erroneous results (n = 700). Next, this method was used to obtain reference values for cardiac function metrics from the UK Biobank. Results: Automated analysis correlated highly with manual analysis for left and right ventricular volumes (all r > 0.95), strain (circumferential r = 0.89, longitudinal r > 0.89), and filling and ejection rates (all r ≥ 0.93). There was no significant bias for cardiac volumes and filling and ejection rates, except for right ventricular end-systolic volume (bias +1.80 ml; p = 0.01). The bias for FT strain was <1.3%. The sensitivity of detection of erroneous output was 95% for volume-derived parameters and 93% for FT strain. Finally, reference values were automatically derived from 2,029 CMR exams in healthy subjects. Conclusions: The study demonstrates a DL-based framework for automated, quality-controlled characterization of cardiac function from cine CMR, without the need for direct clinician oversight.
Article
Full-text available
Objectives: This paper reports the development, validation, and public availability of a new neural network-based system which attempts to identify the manufacturer and even the model group of a pacemaker or defibrillator from a chest radiograph. Background: Medical staff often need to determine the model of a pacemaker or defibrillator (cardiac rhythm device) quickly and accurately. Current approaches involve comparing a device's radiographic appearance with a manual flow chart. Methods: In this study, radiographic images of 1,676 devices, comprising 45 models from 5 manufacturers were extracted. A convolutional neural network was developed to classify the images, using a training set of 1,451 images. The testing set contained an additional 225 images consisting of 5 examples of each model. The network's ability to identify the manufacturer of a device was compared with that of cardiologists, using a published flowchart. Results: The neural network was 99.6% (95% confidence interval [CI]: 97.5% to 100.0%) accurate in identifying the manufacturer of a device from a radiograph and 96.4% (95% CI: 93.1% to 98.5%) accurate in identifying the model group. Among 5 cardiologists who used the flowchart, median identification of manufacturer accuracy was 72.0% (range 62.2% to 88.9%), and model group identification was not possible. The network's ability to identify the manufacturer of the devices was significantly superior to that of all the cardiologists (p < 0.0001 compared with the median human identification; p < 0.0001 compared with the best human identification). Conclusions: A neural network can accurately identify the manufacturer and even model group of a cardiac rhythm device from a radiograph and exceeds human performance. This system may speed up the diagnosis and treatment of patients with cardiac rhythm devices, and it is publicly accessible online.
Article
Full-text available
If anyone doubts the explosive growth of interest in digital medicine, consider a recent conference and workshop in Beijing, jointly organized by the People’s Liberation Army General Hospital and MIT Critical Data to showcase the opportunities and challenges of applying machine learning to the kind of data routinely collected during the provision of care.1 In person, 500 attendees heard a keynote and panels and participated in a health data hackathon. Online, however, the event was streamed to more than one million unique viewers. ©2020 [First paragraph]
Article
Full-text available
Background: Cardiovascular resonance (CMR) imaging is a standard imaging modality for assessing cardiovascular diseases (CVDs), the leading cause of death globally. CMR enables accurate quantification of the cardiac chamber volume, ejection fraction and myocardial mass, providing information for diagnosis and monitoring of CVDs. However, for years, clinicians have been relying on manual approaches for CMR image analysis, which is time consuming and prone to subjective errors. It is a major clinical challenge to automatically derive quantitative and clinically relevant information from CMR images. Methods: Deep neural networks have shown a great potential in image pattern recognition and segmentation for a variety of tasks. Here we demonstrate an automated analysis method for CMR images, which is based on a fully convolutional network (FCN). The network is trained and evaluated on a large-scale dataset from the UK Biobank, consisting of 4,875 subjects with 93,500 pixelwise annotated images. The performance of the method has been evaluated using a number of technical metrics, including the Dice metric, mean contour distance and Hausdorff distance, as well as clinically relevant measures, including left ventricle (LV) end-diastolic volume (LVEDV) and end-systolic volume (LVESV), LV mass (LVM); right ventricle (RV) end-diastolic volume (RVEDV) and end-systolic volume (RVESV). Results: By combining FCN with a large-scale annotated dataset, the proposed automated method achieves a high performance in segmenting the LV and RV on short-axis CMR images and the left atrium (LA) and right atrium (RA) on long-axis CMR images. On a short-axis image test set of 600 subjects, it achieves an average Dice metric of 0.94 for the LV cavity, 0.88 for the LV myocardium and 0.90 for the RV cavity. The mean absolute difference between automated measurement and manual measurement is 6.1 mL for LVEDV, 5.3 mL for LVESV, 6.9 gram for LVM, 8.5 mL for RVEDV and 7.2 mL for RVESV. On long-axis image test sets, the average Dice metric is 0.93 for the LA cavity (2-chamber view), 0.95 for the LA cavity (4-chamber view) and 0.96 for the RA cavity (4-chamber view). The performance is comparable to human inter-observer variability. Conclusions: We show that an automated method achieves a performance on par with human experts in analysing CMR images and deriving clinically relevant measures.
Article
Full-text available
Background Cardiovascular magnetic resonance (CMR) is the gold standard method for the assessment of cardiac structure and function. Reference ranges permit differentiation between normal and pathological states. To date, this study is the largest to provide CMR specific reference ranges for left ventricular, right ventricular, left atrial and right atrial structure and function derived from truly healthy Caucasian adults aged 45–74. Methods Five thousand sixty-five UK Biobank participants underwent CMR using steady-state free precession imaging at 1.5 Tesla. Manual analysis was performed for all four cardiac chambers. Participants with non-Caucasian ethnicity, known cardiovascular disease and other conditions known to affect cardiac chamber size and function were excluded. Remaining participants formed the healthy reference cohort; reference ranges were calculated and were stratified by gender and age (45–54, 55–64, 65–74). ResultsAfter applying exclusion criteria, 804 (16.2%) participants were available for analysis. Left ventricular (LV) volumes were larger in males compared to females for absolute and indexed values. With advancing age, LV volumes were mostly smaller in both sexes. LV ejection fraction was significantly greater in females compared to males (mean ± standard deviation [SD] of 61 ± 5% vs 58 ± 5%) and remained static with age for both genders. In older age groups, LV mass was lower in men, but remained virtually unchanged in women. LV mass was significantly higher in males compared to females (mean ± SD of 53 ± 9 g/m2 vs 42 ± 7 g/m2). Right ventricular (RV) volumes were significantly larger in males compared to females for absolute and indexed values and were smaller with advancing age. RV ejection fraction was higher with increasing age in females only. Left atrial (LA) maximal volume and stroke volume were significantly larger in males compared to females for absolute values but not for indexed values. LA ejection fraction was similar for both sexes. Right atrial (RA) maximal volume was significantly larger in males for both absolute and indexed values, while RA ejection fraction was significantly higher in females. Conclusions We describe age- and sex-specific reference ranges for the left ventricle, right ventricle and atria in the largest validated normal Caucasian population.
Article
Full-text available
Morphological and functional parameters such as chamber size and function, aortic diameters and distensibility, flow and T1 and T2* relaxation time can be assessed and quantified by cardiovascular magnetic resonance (CMR). Knowledge of normal values for quantitative CMR is crucial to interpretation of results and to distinguish normal from disease. In this review, we present normal reference values for morphological and functional CMR parameters of the cardiovascular system based on the peer-reviewed literature and current CMR techniques and sequences. Electronic supplementary material The online version of this article (doi:10.1186/s12968-015-0111-7) contains supplementary material, which is available to authorized users.
Article
Full-text available
Obesity affects a third of adults in the US and results in an increased risk of cardiovascular mortality. While the mechanisms underlying this increased risk are not well understood, animal models of obesity have shown direct effects on the heart such as steatosis and fibrosis, which may affect cardiac function. However, the effect of obesity on cardiac function in animal models is not well-defined. We hypothesized that diet-induced obesity in mice reduces strain, torsion, and synchrony in the left ventricle (LV). Ten 12-week-old C57BL/6 J mice were randomized to a high-fat or low-fat diet. After 5 months on the diet, mice were imaged with a 7 T ClinScan using a cine DENSE protocol. Three short-axis and two long-axis slices were acquired for quantification of strains, torsion and synchrony in the left ventricle. Left ventricular mass was increased by 15% (p = 0.032) with no change in volumes or ejection fraction. Subepicardial strain was lower in the obese mice with a 40% reduction in circumferential strain (p = 0.008) a 53% reduction in radial strain (p = 0.032) and a trend towards a 19% reduction in longitudinal strain (p = 0.056). By contrast, subendocardial strain was modestly reduced in the obese mice in the circumferential direction by 12% (p = 0.028), and no different in the radial (p = 0.690) or longitudinal (p = 0.602) directions. Peak torsion was reduced by 34% (p = 0.028). Synchrony of contraction was also reduced (p = 0.032) with a time delay in the septal-to-lateral direction. Diet-induced obesity reduces left ventricular strains and torsion in mice. Reductions in cardiac strain are mostly limited to the subepicardium, with relative preservation of function in the subendocardium. Diet-induced obesity also leads to reduced synchrony of contraction and hypertrophy in mouse models.
Article
Full-text available
This document is an update to the 2008 publication of the Society for Cardiovascular Magnetic Resonance (SCMR) Board of Trustees Task Force on Standardized Protocols. Since the time of the original publication, 3 additional task forces (Reporting, Post-Processing, and Congenital Heart Disease) have published documents that should be referred to in conjunction with the present document. The section on general principles and techniques has been expanded as more of the techniques common to CMR have been standardized. There is still a great deal of development in the area of tissue characterization/mapping, so these protocols have been in general left as optional. The authors hope that this document continues to standardize and simplify the patient-based approach to clinical CMR. It will be updated at regular intervals as the field of CMR advances.