Left ventricular (LV) dysfunction is mainly assessed by global contractile indices such as ejection fraction and LV Volumes in cardiac MRI. While these indices give information about the presence or not of LV alteration, they are not able to identify the location and the size of such alteration. The aim of this study is to compare the performance of three parametric imaging techniques used in cardiac MRI for the regional quantification of cardiac dysfunction. The proposed approaches were evaluated on 20 patients with myocardial infarction and 20 subjects with normal function. Three parametric images approaches: covariance analysis, parametric images based on Hilbert transform and those based on the monogenic signal were evaluated using cineMRI frames acquired in three planes of views. The results show that parametric images generated from the monogenic signal were superior in term of sensitivity (89.69%), specificity (86.51%) and accuracy (89.06%) to those based on covariance analysis and Hilbert transform in the detection of contractile dysfunction related to myocardial infarction. Therefore, the parametric image based on the monogenic signal is likely to provide additional regional indices about LV dysfunction and it may be used in clinical practice as a tool for the analysis of the myocardial alterations.
Cardiovascular disease is one of the primary causes of death worldwide [
In the recent years, several researchers have focused on the development of new methods for the regional assessment of LV dysfunction. These methods include the automatic detection of endocardial and epicardial myocardial contours using active contour approaches and level set techniques [
The main contributions of this study are:
It compares the performance of three parametric imaging techniques developed by researchers in cardiac MRI for the regional quantification of cardiac dysfunction.
It provides in detail the different steps leading to the computation of three parametric images for the identification of cardiac alterations. This identification is important as some patients with preserved ejection fraction could be diagnosed as healthy based on the use of global indices.
It validates the reproducibility of detecting contractile alterations using three different approaches. Additionally, the comparison between them enable us to choose the most accurate one able to locate the extent of left ventricular dysfunction.
The paper is organized as follows: Section 2 describes the process of parametric imaging computation using three different approaches and the protocol used to compare them. The outcomes of this comparison are presented in Section 3. The obtained results are discussed in Section 4. A summary of this work is presented in Section 5.
This study was carried out in collaboration with the radiology department of Military Hospital of Tunis (HMPIT), Tunisia. Images that used by the proposed study were acquired on a Siemens 3 Tesla MRI scanner (Siemens Medical Solution, Erlangen, Germany) using segmented cineMRI gradient echo sequences with retrospective ECG synchronization. For each patient, cineMRI sequences in two cavities, four cavities and shortaxis views were performed. Each cineMRI sequence consists of 25 frames representing different moments of the cardiac cycle. The following parameters were used for the sequence acquisition: Repetition time (TR) = 3.5 ms; Echo time (TE) = 1.44 ms; thickness = 8 mm; acquisition matrix = 147 * 258.
This retrospective study was performed on a cohort of 40 clinical cases (21 women and 19 men) with an age range from 20 years to 67 years. Among these clinical cases, we distinguish 20 healthy subjects, 20 subjects with myocardial infarction. All variables were presented as the mean ± standard deviation (SD). A linear regression and BlandAltman analysis were used to evaluate the correlation and the degree of agreement between contraction values derived from the different methods. Statistical analysis was performed using the statistical software IBMSPSS Statistics (Windows, version 21.0). The main clinical features of the studied cases are described in
Subjects  Patients (n=20)  Healthy controls (n=20) 

mean ± SD  mean ± SD  
Age (years)  53 ± 15  42 ± 17.2 
LVEF (%)  40.8 ± 15.5  63 ± 4.5 
ESV (ml)  133 ± 69  40 ± 16 
EDV (ml)  207.1 ± 58  141.1 ± 64 
The majority of the parametric imaging approaches described in the literature is evaluated using small clinical populations collected by the authors during their work [
In order to establish an accurate comparison between three parametric methods that have been developed recently for quantification of LV dysfunction, we chose to implement the three methods of parametric images computation following the same validation protocol for the different methods and keeping the same conditions. The different approaches and the steps leading to the computation of different parametric images will be described in the next section.
The parametric imaging technique is based on the measurement of signal variability within the pixel over the cardiac cycle to capture wall motion information. From these extracted signals, it is possible to compute a physical parameter such as the maximum amplitude, the phase or the contraction time calculated for each pixel in the image that reflects the wall motion of the myocardium. All the quantitative parameters were used to generate a map that contains all feature values that reflect the degree of contraction in a welldefined region of interest. The quantitative feature values were projected on a color scale ranging from black to red. If a region suffers from a contraction decrease such as hypokinesia, this will be manifested by a sudden decrease in color indicating the seat of the abnormality [
This technique represents a potential tool in improving the regional detection of wall motion abnormalities. First, the parametric image does not requires the detection of myocardial contours for all images through the cardiac cycle, which is time consuming. Additionally, it provides the exact size of wall motion dysfunction, which manifests by a change of color intensity in the image. Therefore, parametric imaging represents an alternative approach to obtain quantitative indices about motion of underlying structures and to represent an image that summarizes the information of the whole cardiac cycle.
The covariance is a mathematical function for measuring the deviation degree between sets of data that consist of two variables. In medical imaging, this technique was applied to test whether two points or regions in the image move independently or follow the same direction. In cardiology, specifically in cardiac MRI, the covariance was used to measure the deviation of contractility between different myocardial segments in term of wall motion [
The covariance value
The analytic signal is a complex signal computed from the real signal in conjunction with its Hilbert transform [
The Hilbert transform is a phase shift of –Π/2 of the original signal. This analytical representation of the signal allows the simple modeling of the stationary and nonstationary signals. Moreover, some important characteristics of an image could be obtained from the expression of this complex signal, such as the phase and the amplitude, which represent respectively the envelope and the shape of the signal. The instantaneous frequency could also be generated by deriving the phase feature [
For parametric amplitude image computation, it have followed the variation of the pixel intensities through the MRI sequence. A temporal curve was defined for each pixel, reflecting the variation of the LV volume. To compute the analytical signals, this study used the mathematical tool “the Hilbert transform”. This tool makes it possible to add to the real signal an imaginary component. The set of two real and imaginary components forms the complex representation 1D called an analytic representation. From each analytic signal, the maximum value was defined and the instantaneous amplitude feature of the analytical signal was computed using this equation [
This feature has been computed for each pixel within the myocardium. Colorcoding was used to represent all amplitude values as a parametric map (see
The analytical signal has been commonly employed in the field of “image processing” specially, for the extraction of the instantaneous features like amplitude and frequency. By using the mathematical tool Riesz transform as an alternative of the Hilbert transform, it is possible to extend the concept of analytic signal to 2D. This 2D generalization is known as the monogenic signal [
With
In this study, a spherical bandpass filter “logGabor” was used to compute the real component of the monogenic signal [
The process was repeated for all pixel intensity in the MRI image to generate a cartography of amplitudes. The parametric image generated from the monogenic signal gives an information about the contraction level in each region of the myocardium that represents the source of cardiac wall. In this image, a normal left ventricular function is characterized by a homogeneous color, whereas a contractile dysfunction is defined by a variation of color in the myocardial segments. The simplicity of this approach enables its use in the clinical routine of MRI and other imaging methods without any need to draw complex assumptions.
To validate our comparative study, we used the following protocol: for each patient, three parametric amplitude images were computed using three different approaches: covariance analysis, parametric images using Hilbert transform and those based on the monogenic signal. In this study, two experienced cardiologists using the 16 segments model recommended by American Heart Association/American college of cardiology (AHA/ACC) analyzed contraction abnormalities [
After three weeks, two other experienced readers reviewed the same segments to establish a comparison between the three parametric methods and to evaluate the agreement level between each one and the ground truth. The outcomes of this analysis were used to compute sensitivity, specificity, accuracy, positive predictive value and negative predictive value.
Additionally, the average of contraction values generated from the monogenic signal for each myocardial sector were compared to that obtained from the other two approaches. The linear regression and the BlandAltman analysis were computed. For further validation, the execution time for different algorithms were also compared.
All of 640 myocardial segments that use in this study have classified as normal function and dysfunction. Due to the gold standard interpretation, 514 segments suffered from contractile dysfunction (80.31%) while 126 segments had normal function (19.68%). The results of the comparison between the three methods of parametric images computation reveal that the monogenic signalbased approach is more efficient in terms of accuracy (89.06%), sensitivity (89.69%) and specificity (86.51%). The parametric images based on the Hilbert transform is superior in terms of sensitivity (83.07%) and accuracy (82.50%) to the covariance analysis method while this later has a higher specificity (84.92%) than that based on the Hilbert transform (80.16%).
For further evaluation, we opted to evaluate the execution time for different algorithms using 1.73 GHz i7–740QM Processor. According to
As an additional quantitative assessment, the average of contraction values generated from the monogenic signal for each myocardial sector were compared to that obtained from the other two approaches. The linear regression analysis shows a strong correlation between the contraction values obtained by monogenic signal approach and the Hilbert transform with a coefficient of correlation r = 0.91 (for p < 0.01). Furthermore, the BlandAltman analysis shows a mean difference = −2.3 between contraction values derived from the two methods (see
Specificity 
Sensitivity 
Accuracy 
Positive predictive value (PPV)  Negative predictive value (NPV)  

84.92  63.62  67.81  94.11  36.81  
80.16  83.07  82.50  94.47  53.72  
86.51  89.69  89.06  96.44  67.28 
Algorithms  Time (s) 

17  
32  
6.0 
A visual inspection of
Additionally, we compared the findings of our study to the results of other recent methods developed for regional assessment of cardiac wall motion. The results of this comparison are summarized in






SENC Approach 
MR  N=65  Wall motion analysis  Accuracy: 92% 

Strain analysis techniques 
MR  N= 24  Assessment of ventricular dysfunction  Mean difference of peak of strain between the two methods = 1 ± 9% 

2D strain 
MR  N= 126  Evaluation of regional cardiac function  Intraclass correlation coefficient (ICC) for longitudinal strain = 0.55 

Strain analysis 
MR  N=71  Discrimination between different degree of myocardial injury  Accuracy: 78% 

Feature Tracking (FT) techniques 
MR  N= 77  Assessment of myocardial function  Intraobserver agreement: 0.993 

Fast SENC derived Myocardial strain 
MR  N=18  Assessment of myocardial function  ICC for LVEF = 0.92 

Parametric image based on Hilbert transform 
MR  N=20  Regional quantification of LV Function  Pearson's correlation coefficient (r)= 0.983  
Parametric image based on Monogenic signal 
MR  N=40  Regional assessment of cardiac wall motion abnormalities  Accuracy : 89.06% 
The analysis of abnormal myocardial segments using CMRI is usually performed by visual interpretation of cineMRI images and after the computation of some global parameters using manual delineation of the epicardial and endocardial contours [
The outcomes of this study show the approach based on Hilbert transform is superior in terms of sensitivity and accuracy to the covariance analysis method. This result prove the ability of parametric image based on Hilbert transform to correctly identify segments with myocardial dysfunction. Additionally, The Hilbert transform is suitable for both stationary and nonstationary signals but it still suffers from a relatively long acquisition time (23 s/image) that may limit its transfer and accreditation for use in clinical practice. Our findings also demonstrate that the use of monogenic signal show an improvement in the accuracy, specificity and sensitivity of the parametric image interpretation. Thus could be lies to the fact that this approach extracts contraction
In this study, the following limitations should be considered:
The evaluation of three approaches is based the interpretation made by expert radiologists and strain analysis. However, the dependence on the observer skills leads to interobserver variability.
Our comparative study are based on a binary score of normal and abnormal. The classification of myocardial alterations into different categories (moderate hypokinesia, severe hypokinesia, akinesia and dyskinesia) was not performed.
To establish a comparison between three methods, the shortaxis view was only used. Other studies including cineMRI images in two and fourchamber views should be performed.
The described approaches may be adequate for the identification of abnormal segments for patients with myocardial infarction. However, they could be not reproducible for other cardiovascular diseases such as myocarditis. For this reason, other comparative analysis should be conducted and validated on a larger population of clinical cases using more cardiovascular diseases.
In the present study, a comparative study was established between three methods for computing a parametric image using different approaches. The evaluation was based on multiple performance metrics such as accuracy, sensitivity, specificity and execution time. The comparison results shows that the detection performance of contractile alterations using the monogenic signal is superior to those obtained with Hilbert transform and covariance analysis. Additionally, the approach of the monogenic signal offers a rapid acquisition of parametric images and a correct identification of abnormal segments. Hence, this method is likely to give more details about LV dysfunction and it may be used in clinical practice as a tool for the analysis of myocardial dysfunction. A possible way for future research is to evaluate the diagnostic performance of this approach based on the monogenic signal in the study of myocardial viability.
The authors would like to acknowledge University of Tunis el Manar and University of Deusto for providing all facilities and support for this study.