This work presents a fully-automated framework for the pre-processing of free-breathing myocardial perfusion MRI data. Image series are first split into low-rank and sparse components using RPCA. This allows estimation of the deformation fields required to motion correct the image series, in the absence of dynamic contrast enhancement. Once motion corrected, pixels are clustered into anatomically relevant clusters using perfusion-superpixels which groups nearby pixels that have similar time dynamics. A LDA classifier is trained which allows the generation of myocardial probability maps and active contours are fit to the high probability regions to give a delineation of the myocardium.
Motion Correction: The local signal intensity changes caused by the dynamic contrast agent inhibit the application of traditional image registration techniques to retrospectively correct for respiratory motion. To counter-act this, a matrix decomposition technique, robust principal component analysis (RPCA), is employed in order to separate the image series into low-rank and sparse components. It has been found that the low-rank component well models the baseline signal and that the sparse component contains the dynamic signal enhancement (see Figure 1).6 Image registration techniques can then be applied to the low-rank component which contains no dynamic contrast. The computed deformation fields are subsequently used to correct the original image series. The image series motion corrected in such a manner lead to clear and well-defined maximum intensity projections (MIPs), which then allows for significantly more accurate segmentation of the myocardium (see Figure 2).
Segmentation: Pixels are clustered into perfusion-superpixels,7 as shown in Figure 3, using simple linear iterative clustering (SLIC)8 to group pixels that are close in both space and intensity in the MIP, as well as having similar time dynamics. This creates anatomically meaningful groups of pixels that already give an over-segmentation of the myocardium. A feature vector is then created for each superpixel based on the principal components of the time-intensity curves for that superpixel and these feature vectors are used to train a linear discriminant analysis (LDA) classifier. Therefore, given a new unseen image series, the classifier can assign a probability that each of its computed superpixels belong to the myocardium. The probability maps, such as Figure 4, locate the myocardium and active contours can then be fit to give a final segmentation. The result of which is shown in Figure 5.
The motion correction was first evaluated by comparing automatically extracted time-intensity curves before and after motion correction to ground-truth curves. The ground-truth curves are manually extracted using a unique segmentation for each frame whereas the automatically extracted curves use only one segmentation as a mask for the whole images series. The mean normalised mean square error with the ground-truth improves from 0.76 (0.8) to 0.50 (0.62) and the mean Pearson correlation coefficient improves from 0.77 (0.25) to 0.89 (0.17) after motion correction (n=60). It was further shown that there is no statistical difference between pharmacokinetic model parameters computed automatically after this motion correction scheme and the ground-truth values which are obtained through manual motion correction. The average Dice coefficient of overlap between the automatically generated myocardial segmentation and expertly drawn contours is 0.70 using 16 image series, with one segmentation failing to converge. 2-fold cross-validation was employed to prevent the classifier learning the segmentation for a rest series using the stress series of the same patient or vice versa.