Tom Dresselaers1,2, Baptiste Vande Berg1, Sofie Tilborghs3,4, Alexandru Cernicanu5, and Jan Bogaert1,2
1Radiology, UZ Leuven, Leuven, Belgium, 2Dept. of Imaging and Pathology, KU Leuven, Leuven, Belgium, 3Department of Electrical Engineering (ESAT), KU Leuven, Leuven, Belgium, 4Medical Imaging Research Center, UZ Leuven, Leuven, Belgium, 5Philips Benelux, Eindhoven, Netherlands
Synopsis
Keywords: Heart, Radiomics
Several
studies suggest that radiomics of non-contrast cine balanced steady-state free
precession (bSSFP) images allows to detect myocardial infarct patients. Relevant
radiomic features vary however without clear knowledge of the underlying causes.
Most studies rely on time consuming manual contouring on use end-diastolic
images while it was shown that features vary through the heart cycle. In this
study we show that radiomics of end-systolic images result in a higher classification
accuracy than for end-diastolic images and produce the same result when relying
on commercial automated contouring.
Introduction
Recently, the diagnostic potential of radiomics of non-contrast
cine balanced steady-state free precession (bSSFP) cardiac magnetic
resonance (CMR) images has been shown. In particular, the association of radiomics
texture features with acute and chronic myocardial infarction (MI)1-6
and presence of microvascular obstruction7 was demonstrated. Most of these studies utilized end-diastolic
images to train the classification model. Cine bssfp radiomic
features vary however substantially throughout the cardiac cycle in healthy
subjects8. In addition, automated contouring is important in a clinical workflow but the reliability of commercial automated
segmentation tools require further evaluation.
Our first goal was therefore to confirm the feasibility
of distinguishing infarct patients from healthy subjects based on a radiomics
analysis of standard clinical non-contrast mid short axis (SAX) bSSFP cine
images typically acquired during the initial part of a CMR examination and
compare most relevant features with previously reported radiomics models.
Secondly, compare classification models trained using end-systole radiomics features
with end-diastole features. Thirdly, evaluate results derived from automated
contouring using commercial software with manually corrected contours.Methods
Non-contrast midventricular single slice SAX bSSFP
images of 99 consecutive ST-segment elevation acute MI patients and 49 control subjects
were used. In 83 patients (84%) the SAX covered the infarcted myocardium as
visualized at late gadolinium-enhanced CMR. Images were recorded at 1.5T (Ingenia,
Philips) using following parameters: TR=3.7ms, TE=1.7ms, flip angle=60°, accel.
fact. 2, matrix 160 × 256; FOV 300 mm, pixel size 1.6×1.6mm; slice
thickness 8mm, 30 phases). End diastolic (ED) and end systolic (ES) SAX frames
were first automatically segmented (Neosoft suiteHEART®). Next, an in-house
tool (matlab R2019) was used to further exclude residual blood pool pixels (threshold:
mean left ventricular blood pool signal intensity – 1SD). Then, a radiologist blinded to patient type manually
corrected masks when required (3D Slicer9) and scored cine images for
presence of an infarct. Next, radiomic feature extraction was perfomed using PyRadiomics10 (no prior filters; 25 bins). The dataset was then split in a train, validate
and test set (50, 25 and 25%, resp.). Next, the top three features were selected using a fast
correlation based filter (entropy-based measure, which also identifies
redundancy due to pairwise correlations between features). Further selection was based on optimal
classification results (train/validate set) using a support vector machine approach
(Orange vs3.32; linear kernel, cost =1 regression loss epsilon 0.1) applied to the
normalized feature values (µ=0 , σ=1). Finally, classification accuracy (CA) was
assessed on the test data. Pearson correlation coefficients were calculated between ED and ES features
and for the automated vs manually corrected contours.Results
The
three top ranked features for manually contoured ES cine images, RunLengthNonUniformity,
ClusterShade and Median, resulted in a high classification accuracy of 85% on
the test dataset (Fig 2). For manually
contoured ED cine images the top features were RunLengthNonUniformity, Correlation
and Skewness which resulted in a lower CA than the ES model (CA= 77%). As a
reference, the CA of a visual analysis by a radiologist of the ED and ES images
was respectively 65% and 78%. Using the automated
ES myocardial masks as input for the radiomics analysis did not reduce the classification
accuracy and had one identical feature, ClusterShade, and two different
features, SizeZoneNonUniformity and RootMeanSquared (RMS). Both were however strongly
correlated with the top features from the model based on the manually contoured
myocardium (SizeZoneNonUniformity vs. RunLengthNonUniformity r=0.874; RMS vs.
Median:r=0.973 ). In general, correlation between features derived from automated
and manual ES contours was good (40/96 ES features had CC >0.9; figure 3). For
the ED images automated contours resulted in a worse performance (CA=
72%). In general, ED and ES features were poorly correlated with few exception such
as RunLengthNonUniformity (Fig. 4).Discussion
The most
relevant feature to distinguish infarct from healthy subjects was RunLengthNonUniformity.
This feature was previously reported to be important in characterizing the
impact of smoking10 and in differentiation of hypertrophic cardiomyopathy
form controls11. Median was also reported to be an important discriminator
in diabetes suggesting a global alteration of the myocardial tissue10.
A
study including only non-acute MI patients reported first order 10perc and
variance features to be relevant using ED cine images. In our study on acute MI none
of the top feature were correlated with these features (r<0.2) which may be due
to the presence of edema.
It is
yet unclear if partial volume effects, likely to be more relevant in ED than ES,
explain the lower CA based on ED feature. Automated contouring at ED resulted
in a lower CA vs. manual ED contours, which was not the case at ES. This suggests
that ED based radiomics are more sensitive to these partial volumes effects. The reliability of automated contours for radiomics applications needs further validation on a full SAX ES stack of images.Conclusion
Detection
of acute MI using radiomics features of non-contrast mid ventricular SAX cine
images is feasible and more robust relying on ES than ED images. Automated contouring
did not influence the performance of the radiomics model at ES for a mid ventricular slice but needs further evaluation on a full SAX stack. Acknowledgements
No acknowledgement found.References
- Larroza A, López-Lereu MP, Monmeneu JV, et al. Texture
analysis of cardiac cine magnetic resonance imaging to detect nonviable
segments in patients with chronic myocardial infarction. Med Phys.
2018;45(4):1471-1480. doi:10.1002/mp.12783
- Larroza A, Materka A, López-Lereu MP,
Monmeneu JV, Bodí V, Moratal D. Differentiation between acute and chronic
myocardial infarction by means of texture analysis of late gadolinium
enhancement and cine cardiac magnetic resonance imaging. Eur J Radiol.
2017;92:78-83. doi:10.1016/j.ejrad.2017.04.024.
- Nan Zhang, Guang
Yang, Zhifan Gao, Chenchu Xu, Yanping Zhang, Rui Shi, Jennifer Keegan, Lei Xu,
Heye Zhang, Zhanming Fan, and David Firmin; Deep Learning for Diagnosis of Chronic Myocardial Infarction on Nonenhanced Cardiac Cine MRI. Radiology 2019 291:3, 606-617.
- Rauseo E, Izquierdo Morcillo C,
Raisi-Estabragh Z, et al. New Imaging Signatures of Cardiac Alterations in
Ischaemic Heart Disease and Cerebrovascular Disease Using CMR Radiomics. Front
Cardiovasc Med. 2021;8:716577. Published 2021 Sep 23.
doi:10.3389/fcvm.2021.716577
- Avard E, Shiri I, Hajianfar G, et al.
Non-contrast Cine Cardiac Magnetic Resonance image radiomics features and
machine learning algorithms for myocardial infarction detection. Comput Biol
Med. 2022;141:105145. doi:10.1016/j.compbiomed.2021.105145
- Baessler B, Mannil M, Oebel S, Maintz
D, Alkadhi H, Manka R. Subacute and Chronic Left Ventricular Myocardial Scar:
Accuracy of Texture Analysis on Nonenhanced Cine MR Images. Radiology.
2018;286(1):103-112. doi:10.1148/radiol.2017170213
- Peng Y, Wu K, Wang YXJ, Gong J.
Association between cine CMR-based radiomics signature and microvascular
obstruction in patients with ST-segment elevation myocardial infarction. J
Thorac Dis. 2022;14(4):969-978. doi:10.21037/jtd-21-1706
- Alis, D., Yergin, M., Asmakutlu, O.
et al. The influence of cardiac motion on radiomics features: radiomics
features of non-enhanced CMR cine images greatly vary through the cardiac
cycle. Eur Radiol 31, 2706–2715 (2021). https://doi.org/10.1007/s00330-020-07370-y
- 3D
Slicer as an image computing platform for the Quantitative Imaging Network.
Fedorov A, Beichel R, Kalpathy-Cramer J, et al. Magn Reson Imaging.
2012;30:1323–1341.
(12) van
Griethuysen JJM, Fedorov A, Parmar C, et al. Computational radiomics system to
decode the radiographic phenotype. Cancer Res 2017;77(21):e104–e107.
- Cetin I, Raisi-Estabragh Z, Petersen SE, et al. Radiomics Signatures of Cardiovascular Risk Factors in Cardiac MRI: Results From the UK Biobank. Front Cardiovasc Med. 2020;7:591368. Published 2020 Nov 2. doi:10.3389/fcvm.2020.591368
- Baeßler
B, Mannil M, Maintz D, Alkadhi H, Manka R. Texture analysis and machine
learning of non-contrast T1-weighted MR images in patients with hypertrophic
cardiomyopathy-Preliminary results. Eur J Radiol. 2018;102:61-67.
doi:10.1016/j.ejrad.2018.03.013