Empirical Mode Decomposition and Monogenic Signal-Based Approach for Quantification of Myocardial Infarction from MR Images
Abstract: Quantification of myocardial infarction on Late Gadolinium Enhancement cardiovascular magnetic resonance (LGE-CMR) images into heterogeneous infarct periphery (or gray zone) and infarct core plays an important role in cardiac diagnosis, especially in identifying patients at high risk of cardiovascular mortality. However, quantification task is challenging due to noise corrupted in cardiac MR images, the contrast variation, and limited resolution of images. In this study, we propose a novel approach for automatic myocardial infarction quantification, termed DEMPOT, which consists of three key parts: Decomposition of image into intrinsic modes, Monogenic Phase performing on combined dominant modes, and multilevel Otsu Thresholding on the phase. In particular, inspired by the HilbertHuang transform, we perform the multi-dimensional Ensemble Empirical Mode Decomposition and 2D generalization of the Hilbert transform known as the Riesz transform on the MR image to obtain the monogenic phase that is robust to noise and contrast variation. Then, a two-stage algorithm using multilevel Otsu thresholding is accomplished on the monogenic phase to automatically quantify the myocardium into healthy, gray zone, and infarct core regions. Experiments on
LGE-CMR images with myocardial infarction from 82 patients show the superior performance of the proposed approach in terms of reproducibility, robustness, and effectiveness. Existing system:
To assess the myocardial infarction, LGE-CMR imaging is a valuable tool since it provides high spatial resolution and soft tissue discrimination. Recently, studies have shown that in patients with prior myocardial infarction, the extent of gray zone region in myocardium provides valuable information for cardiac diagnosis and is associated with inducibility for ventricular tachycardia (VT) as well as mortality in post- myocardial infarction patients. Therefore, reliable quantification of the myocardial infarction as well as estimation of the infarct size from LGECMR images play an important role and get more attention in medical imaging research field. Proposed system: To quantify the scar regions from LGE cardiac MR images, a various number of approaches have been proposed in the literature, i.e., standard deviation (SD) method. For instance, Kim et al. Defined the scar regions relying on the signal intensity of the remote region in healthy myocardium. Particularly, they first defined a threshold as mean intensity plus two standard deviation of the signal intensity of the remote region. Then, the scar regions are defined as the regions with intensity above the defined threshold. Such approach also is known as SD method and inspires a vast number of researches. Yan et al. Divided the scar regions into infarct core and gray zone parts corresponding to the intensity in the considered regions above three and two standard deviations of the remote region, respectively. Schmidt et al. First calculated the mean, peak, and standard deviation of the remote region. Then, they used a full-width half-maximum (FWHM) criterion to define the scar regions. Tao et al. Combined intensity information and spatial information for extraction and refinement of the scar regions. Advantages:
To further show the advantages when performing monogenic signal on the combined image, we compare the phase when performing monogenic signal on the original input image and the phase when performing monogenic signal on the combined image for some slices in Fig. 2. It is clearly observed from this figure, the disparity between infarcted and healthy regions on the monogenic phases obtained from combined image are more obvious. The proposed DEMPOT method therefore poses some advantages as: First, it does not require selecting a ROI region. Second, it eliminates selecting standard deviation threshold values as in SD approaches. Third, it is more robust and less sensitive to image intensity changes and image contrast variation. Disadvantages: Add a white noise to the data, decompose the white noise-added data into intrinsic mode functions, repeat step with different white noise series each time, and obtain the (ensemble) means of corresponding IMFs of the decompositions as the final result. The ensemble approach has been well-accepted as a solution to the problem of mode mixing. However, the improved capability of Multidimensional EEMD (MEEMD) comes at dramatic increase in the computation cost which has severely limited its application. Modules: Elimination of the ROI Selection: For myocardial infarction quantification, one of well-known approaches is the SD method. In the SD approach, one need to select a ROI in the remote healthy region of the myocardium, then the quantification is based on the mean and standard deviation of the ROI. Due to the need of manual selection of a ROI in each CMR image, the SD approach might not be reproducible for infarction quantification. In addition, the results might be sensitive to the ROI selection, as illustrated in Fig. 5. In this figure, the second row presents the results of the SD method when using the ROI selected by the radiologist, while the third row shows the results when changing the ROI in the remote healthy regions. As can be observed from the second and third rows of Fig. 5, the results of SD method with these two ROIs are
not similar. Unlike the SD method, the proposed DEMPOT method does not require a ROI, thus eliminating the human interactions. The results by DEMPOT method is shown in the fourth row of Fig. 5. As we can see from the second and fourth rows of Fig. 5, the results by DEMPOT method and SD method with ROI selected by radiologist are acceptable and close to those by manual segmentation. Comparison with other quantification approaches: To evaluate the performances of the proposed approach, we compare infarcted regions segmented by DEMPOT and other methods including SD, Level set , Graph cut with the Manual. The results are presented in Fig.6. It can be seen from this figure, compared with other methods, the results obtained by DEMPOT method are close to those by the Manual. The DSCs to measure the similarity between obtained results and the Manual are reported in Table III. As shown in Table III, the DEMPOT method obtained higher DSC values compared to other methods. For further comparison, we apply the SD, Level set, Graph cut and DEMPOT methods for quantification of total infarct regions in all slices from the dataset of 82 patients. The results are also compared with the Manual segmented by the radiologists. The average Dice coefficients of all dataset by the methods are reported in Table IV. As shown in this table, compared to other methods, the proposed DEMPOT approach obtains the highest mean Dice coefficients with smallest standard of deviation (STD) values. Evaluation of the Infarction Area in Predicting Cardiac Mortality in Heart Failure Patients: We apply the SD, Level set, Graph cut, DEMPOT methods, and Manual to quantify infracted regions of all subjects in the data set. Note that the Level set and Graph cut methods as well as Manual are only applied for total infarct quantification. For infarct core and gray zone, the quantification is performed by FWHM approach on the total infarct. The statistical results for all infarction regions are presented in Fig. 9. The results show that the discrimination between survival and cardiac mortality groups is clearly revealed in the gray zone regions quantified by all methods. It is consistent with the hypothesis. Compared to SD, Level set, and Graph cut methods, DEMPOT approach can get smaller p-value, close to those by Manual with FWHM quantification. This demonstrates the higher
discrimination performance between two groups by the proposed DEMPOT. In addition, the ROC of the methods in predicting cardiac death for above methods was shown in Fig. 10. The gray zone region quantified by DEMPOT method can give the best overall discriminative power, compared to SD, Level set, and Graph cut methods. The Robustness of the Proposed Approach: By employing the EEMD and monogenic phase approaches, the proposed DEMPOT method possesses some advantages such as the robustness to noise and illumination changes. It is demonstrated that the local phase achieved from EEMD and monogenic signal can be used as the image representation for infarction quantification in cardiac MR images. Due to the clear disparity between the healthy and infarcted regions in local phase, a multilevel Otsu thresholding approach is applicable to the local phase image, rather than the original CMR image. This is obvious when the intensity in myocardium region of the CMR image is with narrow range or image histogram in the myocardium region is not clearly distinguished. In addition by using an automatic approach based on multilevel Otsu thresholding for finding optimal thresholds, our method does not require selecting suitable standard deviation values, such as those ranging from 2 to 6 as . Besides, the method eliminates the human interaction for selecting the myocardium ROI in each CMR image as in SD approaches, such as in Yan et al., thus allows reproducibility for myocardial infarction.