2) MYOCARDIAL SEGMENTATION FROM PERFUSION IMAGES

The myocardium is detected on each slice location sequentially starting from the basal and proceeding to the apical slices. Figure 1(​(aa–c) outlines example result of the process, while the individual steps are more precisely outlined in Figure 2. In the basal slice, the LV boundary, Figure 1(a), is refined using region growing from the initial LV cavity mask using the frame of peak LV intensity. A convex hull is applied also ensure that any the papillary muscles present are included with in the LV boundary. Selecting the enhanced LV boundary as the endocardial boundary ensures the exclusion of the papillary muscles. The epicardial edge of the myocardium is detected from a time-signal intensity normalized image series. This is distinct from 2D image normalization which, in general, sets the maximum intensity value to 1, the minimum to 0, and appropriately scales the intermediate values to the same range. Instead, we normalize each pixel individually based on its intensity range over time. That is, the pixels’ maximum value over the time series is set to 1, while its minimum is set to 0, and intermediate values are scaled appropriately to this range. This pixel and time-based normalization highlights the relative contrast enhancement timing of each pixel rather than its absolute signal intensity magnitude. This is useful in cases with perfusion defects which do not enhance with contrast as much as healthy tissue. The normalization amplifies this small enhancement to the same level as healthy myocardial enhancement to provide better contrast for segmentation. Using this normalized series, a rough estimate of the edge is detected from a baseline intensity image, in which the myocardium and LV are dark while the background tissue is relatively bright. We chose the baseline image because there is generally a larger relative intensity difference between the mid-level background and the dark baseline myocardium than there is to the mid-level background and the perfusing myocardium. This image is transformed into the polar domain centered on the LV, and Canny edge detection is applied [19]. The longest continuous edge is extended around the myocardium using a polynomial fitting and selected as the initial epicardial edge and is transformed back to the cartesian domain. Finally, this edge is refined using an active contours algorithm [20] performed on a contrast enhanced image during the washout phase. This image is reconstructed using principal component analysis (PCA) [21] to remove redundant information and extract the most prevalent myocardial contrast enhancement information. We construct the principal component images using the images after the peak contrast enhancement (i.e. during the washout phase) and select the first principal component image that contains the primary information of the images. This first principal component image has the effect of noise reduction and signal intensity contrast improvement for better myocardial boundary detection.

An outline of the automated image processing steps for CMR perfusion image segmentation. This is part of the proposed automated pixel-wise MBF quantification processing pipeline as shown in Figure 1.

The same process is applied to the mid and apical slices, with one extra step: the LV and RV must be re-located by cross correlating the heart region pixels with the ventricles signal measured from the previous slice. The pixel with the highest coefficient of correlation is selected as a seed point, from which the remainder of the ventricle is region grown; this step is similar to the LV refinement performed on the basal slice. The LV in all slices is region grown based on the peak intensity frame as measured from the basal slice. It is possible that the ventricles, particularly the RV, may not be visible in the apical slice. If the RV is not detected, the processing continues without it. However, the processing for myocardial segmentation is aborted if the LV cannot be located.