TECHNICAL DEVELOPMENT

Automated Inline Analysis of Myocardial Perfusion MRI

with Deep Learning
Hui Xue, PhD • Rhodri H. Davies, MD, PhD • Louise A. E. Brown, MD, PhD • Kristopher D. Knott, MD, MBBS •
Tushar Kotecha, MD, PhD • Marianna Fontana, MD, PhD • Sven Plein, MD, PhD • James C. Moon, MD •
Peter Kellman, PhD
From the National Heart, Lung and Blood Institute, National Institutes of Health, 10 Center Dr, Bethesda, MD 20892 (H.X., P.K.); Barts Heart Centre, Barts Health NHS
Trust, London, England (R.H.D., K.D.K., J.C.M.); Department of Biomedical Imaging Science, Leeds Institute of Cardiovascular and Metabolic Medicine, University of
Leeds, Leeds, England (L.A.E.B., S.P.); and National Amyloidosis Centre, Royal Free Hospital, London, England (T.K., M.F.). Received January 28, 2020; revision requested
February 24; revision received June 1; accepted June 16. Address correspondence to H.X. (e-mail: [email protected]).

Conflicts of interest are listed at the end of this article.

Radiology: Artificial Intelligence 2020; 2(6):e200009 • https://doi.org/10.1148/ryai.2020200009 • Content codes:

Purpose: To develop a deep neural network–based computational workflow for inline myocardial perfusion analysis that automatically
delineates the myocardium, which improves the clinical workflow and offers a “one-click” solution.

Materials and Methods: In this retrospective study, consecutive adenosine stress and rest perfusion scans were acquired from three hospitals
between October 1, 2018 and February 27, 2019. The training and validation set included 1825 perfusion series from 1034 patients
(mean age, 60.6 years 6 14.2 [standard deviation]). The independent test set included 200 scans from 105 patients (mean age, 59.1
years 6 12.5). A convolutional neural network (CNN) model was trained to segment the left ventricular cavity, myocardium, and
right ventricle by processing an incoming time series of perfusion images. Model outputs were compared with manual ground truth for
accuracy of segmentation and flow measures derived on a global and per-sector basis with t test performed for statistical significance.
The trained models were integrated onto MR scanners for effective inference.

Results: The mean Dice ratio of automatic and manual segmentation was 0.93 6 0.04. The CNN performed similarly to manual
segmentation and flow measures for mean stress myocardial blood flow (MBF; 2.25 mL/min/g 6 0.59 vs 2.24 mL/min/g 6 0.59, P =
.94) and mean rest MBF (1.08 mL/min/g 6 0.23 vs 1.07 mL/min/g 6 0.23, P = .83). The per-sector MBF values showed no differ-
ence between the CNN and manual assessment (P = .92). A central processing unit–based model inference on the MR scanner took
less than 1 second for a typical perfusion scan of three slices.

Conclusion: The described CNN was capable of cardiac perfusion mapping and integrated an automated inline implementation on the
MR scanner, enabling one-click analysis and reporting in a manner comparable to manual assessment.

Supplemental material is available for this article.

Published under a CC BY 4.0 license.

M yocardial perfusion MRI has proven to be an accurate,

noninvasive imaging technique to detect ischemic heart
disease (1). Quantitative MR perfusion is more objective (2),
In this study, we proposed a deep CNN-based com-
putational workflow for myocardial perfusion analysis
using MRI. The right ventricular (RV) insertion points
and automated in-line methods (3,4) offer improved effi- were determined to allow reporting of perfusion accord-
ciency of analysis. Compared with qualitative visual assess- ing to the standard 16-segment model proposed by the
ment, quantitative methods improve the detection of disease American Heart Association (AHA). To use the dynamic
with a global reduction in flow as seen in balanced multives- change of intensity that was the result of contrast material
sel obstruction or microvascular disease (5). uptake, the proposed solution operates on the time series
Without automated segmentation of the MR perfu- of two-dimensional (2D) perfusion images (referred to
sion maps, a reporting clinician would have to manually here as 2D+T) after respiratory motion correction. The
draw regions of interest to extract global or regional flow performance of the trained CNNs was quantitatively
values. Objective perfusion assessment could be further evaluated by comparing against manually established
facilitated by segmenting the myocardium to automati- ground truth for both segmentation accuracy and global
cally generate the report, leading to a one-click solution as well as regional flow measures on an independent hold-
to improve workflow. Automated MR perfusion mea- out test dataset.
surement could serve as the input for downstream car- To promote the clinical validation and adoption of the
diovascular disease classification (6) in which pretrained proposed solution, the trained deep learning models were
convolutional neural network (CNN) models receive integrated onto MR scanners using the Gadgetron Inlin-
myocardial flow and other imaging features to predict eAI toolbox (7). The CNN models were applied to the ac-
the probability of ischemic heart disease. These previous quired images as part of the scanner computing workflow
studies used manual segmentation and can be automated (inline processing) at the time of the scan, rather than as a
with the proposed approach. part of postprocessing. The resulting segmentation results

This copy is for personal use only. To order printed copies, contact [email protected]
Automated Inline Analysis of Myocardial Perfusion MRI

between the training and validation data to the independent

Abbreviations test data (10). Among the assembled independent test data,
AHA = American Heart Association, CI = confidence interval, 96 scans were acquired from 3-T scanners and 104 were from
CNN = convolutional neural network, LV = left ventricle, MBF =
myocardial blood flow, RV = right ventricle, 2D = two-dimensional 1.5-T scanners.

Summary MRI Acquisition

The described convolutional neural network was capable of segment-
ing and determining the mean stress and rest myocardial blood flow
Perfusion imaging used a previously published dual-sequence
in a manner comparable to manual segmentation. scheme (3). The imaging started after administering the con-
trast agent and with the acquisition of typical three 2D im-
Key Points ages cutting through the heart. This acquisition was repeated
n The study proposed and validated a convolutional neural network for every heartbeat to capture the contrast material passage.
solution for cardiac perfusion mapping and integrated an auto- Details for MRI can be found in Appendix E1 (supplement).
mated inline implementation on the MR scanner, enabling one-
click analysis and reporting.
n The large training set included 1825 perfusion series from 1034 Data Preparation and Ground Truth Labeling
patients (mean age, 60.6 years 6 14.2), and the independent test The perfusion image series underwent motion correction and
set included 200 scans from 105 patients (mean age, 59.1 years 6 surface coil inhomogeneity correction. Resulting images were
12.5).
spatially upsampled to 1.0 mm2, and the central field of view
n Comparison of automated- and manual-derived myocardial blood
flow measurement showed no differences on both global and per-
(176 3 176 mm2) was cropped. For the short-axis perfusion
sector basis (P . .80). slices, the LV endo- and epicardial boundaries were manually
traced, together with the RV (Fig 1). Information for data prepa-
ration and labeling can be found in Appendix E2 (supplement).
and analysis reports were available for immediate evaluation
prior to the next image series. The method described here has Model and Training
been used in a prospective study of more than 1000 patients The U-Net semantic segmentation architecture (11,12) was
to demonstrate the prognostic significance of quantitative stress adopted for the perfusion segmentation. The neural net (Fig
perfusion (8). A one-click solution to acquire free-breathing per- 2) consisted of downsampling and upsampling layers, each
fusion images, perform pixelwise flow mapping, and conduct including a number of ResNet blocks (13) with batch nor-
automated analysis with a 16-segment AHA report generated on malization (14) and leaky rectified linear unit (15) nonlinear-
the MR scanner was demonstrated. ity. The data for training were split into a training set (87.5%
of all studies) and a validation set (12.5% of all studies). The
Materials and Methods CNN model and optimization were implemented using Py-
Torch (16).
Imaging and Data Collection The trained model was integrated to run on MR scan-
In this retrospective study, the datasets consisted of adenosine ners using the Gadgetron Inline AI (7) streaming software.
stress and rest perfusion scans which were acquired at three A screenshot (Fig 3) illustrates the perfusion mapping with
hospitals (Barts Heart Centre, BHC; Royal Free Hospital, overlaid CNN-based segmentation and AHA report applied
RFH; Leeds Teaching Hospitals, LTH) between October 1, to a patient with reduced regional perfusion. This is a one-
2018 and February 27, 2019. Data were acquired with the re- click solution for automated analysis of quantitative perfu-
quired ethical and/or audit secondary use approvals or guide- sion flow mapping.
lines (as per each center) that permitted retrospective analysis Appendix E3 (supplement) provides details about model,
of anonymized data for the purpose of technical development, training, and inline integration.
protocol optimization, and quality control. All data were ano-
nymized and delinked for analysis by the National Institutes Statistical Analysis
of Health with approval by the National Institutes of Health The segmentation of automated processing was compared
Office of Human Subjects Research (exemption #13156). The with the manually labeled test set. Performance was quanti-
collected datasets were previously included in a recent study fied in both segmentation accuracy and myocardial flow mea-
(9) that developed a left ventricular (LV) blood pool detection sures. The Dice ratio for manual label and automatic segmen-
solution for arterial input function images, whereas this study tation masks was computed, together with the false-positive
used the datasets for perfusion myocardium segmentation. and false-negative errors. A false-positive error was defined
A total of 1825 perfusion scans from 1034 patients (mean as the percentage area of the segmented mask in the CNN
age, 60.6 years 6 14.2; 692 men) were assembled and split result that was not labeled in the manual one. A false-negative
into training and validation sets used for CNN model train- error was defined as the percentage area of segmented mask
ing. An independent hold-out consecutive test set was as- in the manual that was not labeled in the automated result.
sembled, consisting of 200 perfusion scans from 105 patients The precision (defined as the percentage of segmented area
(mean age, 59.1 years 6 12.5; 76 men). The Table sum- in both the CNN and manual masks divided by CNN area)
marizes detailed dataset information. There was no overlap and recall (defined as the percentage of segmented area in

2radiology-ai.rsna.org n Radiology: Artificial Intelligence Volume 2: Number 6—2020

 Xue et al

Dataset Information Results

Dataset and Site No. of Patients No. of Stress Scans No. of Rest Scans CNN Segmentation Overview and Optimization
Training An example of segmentation (Fig 4) illustrates the
BHC 475 432 475 contours overlaid on perfusion images and corre-
RFH 345 219 345 sponding flow maps. The trained CNN correctly
LTH 214 140 214 delineated the LV cavity and myocardium. The RV
Total 1034 791 1034 insertion direction was accurately detected to allow
Test sector division. The epicardial fat was correctly ex-
BHC 54 46 54 cluded from segmentation and the papillary muscles
RFH 13 11 13 were avoided.
LTH 38 38 38
Total 105 95 105 CNN and Ground Truth Segmentation
Performance Comparisons
The mean Dice ratio of myo-
cardium segmentation between
CNN and manual ground truth
was 0.93 6 0.04 (90% confi-
dence interval [CI]: 0.88, 0.97).
False-positive and false-negative
rates were 0.09 6 0.06 (90%
CI: 0.02, 0.18) and 0.06 6 0.05
(90% CI: 0.005, 0.13). Preci-
sion and recall were 0.92 6 0.06
(90% CI: 0.81, 0.97) and 0.94
Figure 1: Data preparation for performing convolutional neural network (CNN)–based segmentation used in this study. 6 0.05 (90% CI: 0.87, 0.99).
Respiratory motion correction of perfusion images provides pixelwise alignment of myocardial tissue. Image intensities are cor- The myocardium boundary er-
rected for surface coil inhomogeneity and converted to gadolinium (Gd) concentration units. Images are resampled to a fixed ror was 0.33 mm 6 0.15 (90%
temporal and spatial resolution and cropped around the left ventricular cavity. The resulting 2D+T time series of images is input CI: 0.13 mm, 0.52 mm). Given
for CNN training, together with supplied manual labeling. 2D+T = two-dimensional perfusion images.
the training image spatial resolu-
tion of 1 mm2, the mean bound-
ary error was less than 0.5 pixels.
both the CNN and manual masks divided by manual area) The mean bidirectional Hausdorff distance was 2.52 mm 6
were also reported. The myocardium boundary errors (17), 1.08 (90% CI: 1.42 mm, 4.13 mm), and the mean angle be-
defined as the mean distance between myocardial borders of tween auto and manually determined RV insertion point direc-
two masks, and the Hausdorff distance (18) were computed tions was 2.65° 6 3.89 (90% CI: 0.28°, 5.95°).
for the endo- and epicardium borders. The detection accu- The mean stress flow was 2.25 mL/min/g 6 0.59 for the
racy of RV insertion was measured by the angular difference CNN and 2.24 mL/min/g 6 0.59 for manual segmentation (P
between auto- and manual-determined direction vectors for = .94). For rest scans, the CNN gave 1.08 mL/min/g 6 0.23 and
RV insertion because only the orientation was needed for seg- manual measure gave 1.07 mL/min/g 6 0.23 (P = .83). The per-
mentation. Global and per-sector myocardial flow measures sector measures showed no difference between the CNN and
were used to quantify the CNN performance compared with manual measures (P = .92). Bland-Altman plots (Fig 5) compare
manual results, displayed using Bland-Altman plots. In addi- automatic to manual processing of MBF for both global MBF
tion, contours were visually inspected for segmentation fail- and 16-sector values.
ures on all 200 test scans. The performance was further evaluated separately for 3-T
Results were presented as mean 6 standard deviation. A t and 1.5-T test scans. The mean Dice ratio was 0.93 6 0.04 for
test was performed, and a P value less than .05 was considered 3 T and 0.93 6 0.03 for 1.5 T (P = .97). At 3 T, the mean stress
statistically significant (Matlab R2017b; MathWorks, Natick, flow was 2.20 mL/min/g 6 0.59 for CNN and 2.21 mL/min/g
Mass). A t test was used to test whether there were significant 6 0.59 for manual (P = .93). The mean rest flow was 1.08 mL/
differences on myocardial blood flow (MBF) values derived min/g 6 0.23 for CNN and 1.07 mL/min/g 6 0.23 for manual
from manual and automated segmentation. (P = .84). At 1.5 T, the mean stress flow was 2.29 mL/min/g 6
0.60 for the CNN and 2.29 mL/min/g 6 0.59 for manual (P =
CNN Data Sharing .97). The mean rest flow was 1.08 mL/min/g 6 0.23 for CNN
To encourage researchers on other platforms to adopt the pro- and 1.07 mL/min/g 6 0.23 for manual (P = .93).
posed solution, the CNN model files and other resources are Contours were visually evaluated on all 200 test scans (three
shared openly (https://github.com/xueh2/QPerf). slices each). In a single stress case, one slice failed to properly

Radiology: Artificial Intelligence Volume 2: Number 6—2020 n radiology-ai.rsna.org 3

Automated Inline Analysis of Myocardial Perfusion MRI

Figure 2: Schematic plot of the convolutional neural network (CNN) trained in this study. This network consists of downs-
ampling and upsampling layers. Each layer includes a number of ResNet blocks. More layers and blocks can be inserted
into the CNN to increase its depth. In the example illustration, two layers are used with two blocks for each layer. The total
number of convolution blocks is 23. BN = batch normalization, C = concatenate filter response, CONV = convolution, LV =
left ventricle, RELU = rectified linear units, RV = right ventricle.

segment the RV; however, the myocardium was properly seg- proximately 250 msec per slice. For a typical three short-axis
mented. A second rest case had one apical slice in which the acquisition, inline analysis took less than 1 second on the cen-
myocardium segmentation included blood pool. There was ap- tral processing unit. On the older Xeon E5, model loading
parent through-plane motion that was uncorrected. No other time was approximately 130 msec and applying models took
segmentation failures were found. 370 msec per slice.

CNN Speed Performance Discussion

The CNN model was integrated on the MR scanner. On Xeon This study presented a deep neural network–based workflow
Gold, model loading time was approximately 120 msec, and for automated myocardial segmentation and reporting of the
the applying model on the incoming perfusion series was ap- AHA 16-sector model for pixelwise perfusion mapping. The

4radiology-ai.rsna.org n Radiology: Artificial Intelligence Volume 2: Number 6—2020

 Xue et al

Figure 3: Screenshot in a patient undergoing an adenosine stress study demonstrates the proposed inline analysis solution on an MR
scanner. Stress maps show regional flow reduction in septal and inferior sectors. The determined right ventricular insertion was used to split
myocardium to American Heart Association (AHA) sectors, with the contours overlaid to mark territories. The inline reporting further produced
a 16-sector AHA bulls-eye plot with global and per-sector flow measures reported in a table.

derived myocardial measures were computed and reported and myocardial perfusion reserve measured automatically using
inline on the MR scanner, taking just 1 additional second of artificial intelligence quantification of cardiac MR perfusion (2).
inline processing time. Quantitative evaluation in this initial This study demonstrates the relevance of automated myocardial
study demonstrated performance of myocardial segmentation segmentation in cardiac MR stress perfusion.
and sector-based analysis that was well matched to a human Automated cardiac MR image analysis has been attempted
expert. This study used stress and rest data from seven scanners over a long period of time (21). Most work focused on cine im-
at three sites at two field strengths, using more than 1800 con- age analysis (eg, MICCAI 2017 ACDC challenge [22]). The first
secutive scans for training and 200 for test. Bland-Altman anal- deep learning study, which was based on a large data cohort and
ysis demonstrated a 95% CI for global MBF of 0.05 mmol/ reported performance matching human level, was published in
min/g compared with manual labeling, which was sufficient for 2018 (23). Since then, deep neural nets were applied to other
automated detection and reporting. Prior work on segment- cardiac MR imaging applications, such as T1 mapping and car-
ing perfusion (19) used much smaller datasets that resulted in diac late enhancement segmentation (24). Our approach used
much higher variance. A weighted sum loss function was used the temporal information through the whole bolus passage
in this study and gave good accuracy. There are indeed many (2D+T) to exploit the contrast dynamics for detecting both RV
other alternatives, such as soft Dice ratio or focal loss (20), and LV which enabled finding the RV insertion point. The epi-
which can be effective in perfusion segmentation task. Which cardial fat, showing no dynamic intensity changes, was correctly
loss function is the best may vary for different applications. excluded from segmentation which would have been more dif-
A comprehensive overview and implementation of many loss ficult to avoid on a single static image.
functions can be found at https://github.com/JunMa11/SegLoss. This study had several limitations. First, the presented study
The prognostic significance of proposed artificial intelligence was conducted on MR scanners from a single vendor. Although
application was studied and recently published in Knott et al (8). the specific imaging dual sequence used may not be available on
In this study, 1049 patients with known or suspected coronary other platforms, the proposed segmentation method and CNN
artery disease underwent stress MR perfusion scans and were an- models may be still applicable. Second, although the proposed
alyzed with the proposed CNN models, showing reduced MBF algorithm works well for most cases, a few cases are challenging.

Radiology: Artificial Intelligence Volume 2: Number 6—2020 n radiology-ai.rsna.org 5

Automated Inline Analysis of Myocardial Perfusion MRI

Figure 4: Adenosine stress perfusion images and myocardial blood flow (MBF) maps illustrate segmentation in the format of derived American
Heart Association sector contours overlaid on flow maps. For each case, the first row is the images in gadolinium units and the second row is the MBF
maps. Sector contours were overlaid to mark three territories for left anterior coronary artery (yellow), right coronary artery (green), and left circumflex
(red). (a) Patient with single-vessel obstructive coronary artery disease in right coronary artery territory. Papillary muscle was not included in segmenta-
tion. (b) Patient with hypertrophic cardiomyopathy illustrates that the convolutional neural network–based segmentation works with thick myocardium
and small cavity. The epicardial fat was correctly excluded.

For example, in the case of severe hypertrophy, some slices may extend segmentation to long-axis slices. In the case where the
not exhibit any blood pool (ie, complete extinction) where no basal slice may cover some portion of the outflow tract, the
endocardial contours are drawn. Third, the CNN models are proposed algorithm will avoid the blood pool and divide a seg-
currently trained for short-axis slices and cannot be applied to ment accordingly or may skip a segment entirely, resulting in
long-axis views. New training and test datasets are needed to incomplete segmentation. Fourth, in cases of severe respiratory

6radiology-ai.rsna.org n Radiology: Artificial Intelligence Volume 2: Number 6—2020

 Xue et al

Figure 5: Bland-Altman plots for independent test dataset (a) global mean myocardial blood flow (MBF) and (b) per-
sector measures. No significant differences were found between convolutional neural network (CNN)–derived results and
manual measures. The dotted lines mark the 95% confidence range.

motion that is beyond the capacity of the in-plane retrospective sure any questions related to the work are appropriately resolved, all authors; litera-
ture research, H.X., M.F., P.K.; clinical studies, L.A.E.B., K.D.K., T.K., M.F., S.P.,
methods, portions of the myocardium may be blurred, where J.C.M.; experimental studies, H.X., R.H.D., P.K.; statistical analysis, H.X., P.K.;
CNN segmentation can perform poorly. However, in these cases and manuscript editing, H.X., R.H.D., L.A.E.B., K.D.K., M.F., S.P., J.C.M., P.K.
manual segmentation is difficult as well.
Another limitation of this study was the single operator for data Disclosures of Conflicts of Interest: H.X. disclosed no relevant relationships. R.H.D.
disclosed no relevant relationships. L.A.E.B. disclosed no relevant relationships. K.D.K.
labeling. The interoperator reliability was not tested in this article. disclosed no relevant relationships. T.K. disclosed no relevant relationships. M.F. dis-
Because this solution had been deployed to MR scanners, clinical closed no relevant relationships. S.P. disclosed no relevant relationships. J.C.M. Activi-
collaborators have started to use this solution (8,25), but more ties related to the present article: institution receives grants from British Heart Founda-
tion for research fellows to develop perfusion mapping as a clinical tool. Activities not
clinical validation is required to further validate this solution. related to the present article: disclosed no relevant relationships. Other relationships:
In conclusion, we demonstrated automated analysis can be disclosed no relevant relationships. P.K. disclosed no relevant relationships.
achieved on clinical scanners for perfusion MRI. Deep learning
enabled inline analysis immediately after data acquisition as part References
of imaging computation, therefore making the process more ob- 1. Nagel E, Greenwood JP, McCann GP, et al. Magnetic resonance perfusion or
jective, convenient, and faster, reducing clinical burden. fractional flow reserve in coronary disease. N Engl J Med 2019;380(25):2418–
2428.
2. Kotecha T, Martinez-Naharro A, Boldrini M, et al. Automated Pixel-Wise
Author contributions: Guarantors of integrity of entire study, H.X., M.F.; study Quantitative Myocardial Perfusion Mapping by CMR to Detect Obstruc-
concepts/study design or data acquisition or data analysis/interpretation, all authors; tive Coronary Artery Disease and Coronary Microvascular Dysfunction:
manuscript drafting or manuscript revision for important intellectual content, all Validation Against Invasive Coronary Physiology. JACC Cardiovasc Imaging
authors; approval of final version of submitted manuscript, all authors; agrees to en- 2019;12(10):1958–1969.

Radiology: Artificial Intelligence Volume 2: Number 6—2020 n radiology-ai.rsna.org 7

Automated Inline Analysis of Myocardial Perfusion MRI

3. Kellman P, Hansen MS, Nielles-Vallespin S, et al. Myocardial perfusion International Conference on Machine Learning (ICML). Volume 37. Lillie,
cardiovascular magnetic resonance: optimized dual sequence and reconstruc- France: 2015;448–456.
tion for quantification. J Cardiovasc Magn Reson 2017;19(1):43. 15. Maas AL, Hannun AY, Ng AY. Rectifier nonlinearities improve neural
4. Xue H, Brown LAE, Nielles-Vallespin S, Plein S, Kellman P. Automatic network acoustic models. Proceedings of the 30th International Conference
in-line quantitative myocardial perfusion mapping: Processing algorithm on Machine Learning (ICML). Volume 28. Atlanta, Ga: 2013.
and implementation. Magn Reson Med 2020;83(2):712–730. 16. Paszke A, Gross S, Chintala S, et al. Automatic differentiation in pytorch.
5. Zorach B, Shaw PW, Bourque J, et al. Quantitative cardiovascular magnetic 31st Conf Neural Inf Process Syst, 2017; 1–8.
resonance perfusion imaging identifies reduced flow reserve in microvascular 17. Xue H, Shah S, Greiser A, et al. Motion correction for myocardial T1 map-
coronary artery disease. J Cardiovasc Magn Reson 2018;20(1):14. ping using image registration with synthetic image estimation. Magn Reson
6. Acampa W, Petretta M, Evangelista L, et al. Myocardial perfusion imaging Med 2012;67(6):1644–1655.
and risk classification for coronary heart disease in diabetic patients. The 18. Huttenlocher DP, Rucklidge WJ, Klanderman GA. Comparing images
IDIS study: a prospective, multicentre trial. Eur J Nucl Med Mol Imaging using the Hausdorff distance under translation. Proceedings 1992 IEEE
2012;39(3):387–395. Computer Society Conference on Computer Vision and Pattern Recognition,
7. Xue H, Davies R, Hansen D, et al. Gadgetron Inline AI: Effective Model Champaign, IL, June 15–18, 1992. Piscataway, NJ: IEEE, 1992; 654–656.
inference on MR scanner [abstr]. In: Proceedings of the Twenty-Seventh 19. Scannell CM, Veta M, Villa ADM, et al. Deep-Learning-Based Preprocess-
Meeting of the International Society for Magnetic Resonance in Medicine. ing for Quantitative Myocardial Perfusion MRI. J Magn Reson Imaging
Berkeley, Calif: International Society for Magnetic Resonance in Medicine, 2020;51(6):1689–1696.
2019; 2019; 4837. 20. Lin TY, Goyal P, Girshick R, He K, Dollar P. Focal Loss for Dense Object
8. Knott KD, Seraphim A, Augusto JB, et al. The Prognostic Significance of Detection. IEEE Trans Pattern Anal Mach Intell 2020;42(2):318–327.
Quantitative Myocardial Perfusion: An Artificial Intelligence-Based Approach 21. Peng P, Lekadir K, Gooya A, Shao L, Petersen SE, Frangi AF. A review of
Using Perfusion Mapping. Circulation 2020;141(16):1282–1291. heart chamber segmentation for structural and functional analysis using
9. Xue H, Tseng E, Knott KD, et al. Automated detection of left ventricle cardiac magnetic resonance imaging. MAGMA 2016;29(2):155–195.
in arterial input function images for inline perfusion mapping using deep 22. Bernard O, Lalande A, Zotti C, et al. Deep Learning Techniques for Auto-
learning: A study of 15,000 patients. Magn Reson Med 2020. 10.1002/ matic MRI Cardiac Multi-Structures Segmentation and Diagnosis: Is the
mrm.28291. Published online May 7, 2020. Problem Solved? IEEE Trans Med Imaging 2018;37(11):2514–2525.
10. Bluemke DA, Moy L, Bredella MA, et al. Assessing radiology research on 23. Bai W, Sinclair M, Tarroni G, et al. Automated cardiovascular magnetic
artificial intelligence: A brief guide for authors, reviewers, and readers-from resonance image analysis with fully convolutional networks. J Cardiovasc
the Radiology Editorial Board. Radiology 2020;294(3):487–489. Magn Reson 2018;20(1):65.
11. Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for bio- 24. Moccia S, Banali R, Martini C, et al. Development and testing of a deep
medical image segmentation. Lect Notes Comput Sci 2015;9351:234–241. learning-based strategy for scar segmentation on CMR-LGE images. MAGMA
12. Zhang Z, Liu Q, Wang Y. Road Extraction by Deep Residual U-Net. IEEE 2019;32(2):187–195.
Geosci Remote Sens Lett 2018;15(5):749–753. 25. Camaioni C, Knott KD, Augusto JB, et al. Inline perfusion mapping provides
13. Xie S, Girshick R, Dollár P, Tu Z, He K. Aggregated residual transformations insights into the disease mechanism in hypertrophic cardiomyopathy. Heart
for deep neural networks. 2017 IEEE Conference on Computer Vision and 2020;106(11):824–829.
Pattern Recognition (CVPR), Honolulu, HI, July 21–26, 2017. Piscataway,
NJ: IEEE, 2017; 5987–5995.
14. Ioffe S, Szegedy C. Batch Normalization: Accelerating Deep Network
Training by Reducing Internal Covariate Shift. Proceedings of the 32nd

8radiology-ai.rsna.org n Radiology: Artificial Intelligence Volume 2: Number 6—2020