One identifying characteristic of damaged myocardium (heart muscle)  is reduced contraction during systole. In particular, the wall motion of the left ventricle of the heart is of great concern as this is where most heart Problems occur. However, this is a difficult problem because while many images can be taken of the left ventricle, it is practically impossible for a medical expert to extract not only anatomical and physiological information, but also temporal data from two hundred plus  images.

Heart image shortly after tag planes are applied and after contraction

Horizontally tagged short-axis, vertically tagged short axis, and long axis

        Having access to a set of tagged MRI images, Todd Delaune and I would like to attempt to create a very basic visualization of the left ventricle for our vision project. Our primary impetus will focus on tracking the motion of the myocardium in order to construct a three-dimensional wire model of the left ventricle – varying over time.  We have created two goals depending on how far we get into the project and on available time. The first goal will be creating a surface “wire frame” that corresponds with the outer wall motion of the left ventricle.  Should everything run smoothly and as planned, we would like to take advantage of the tagged images and extend the first goal so that we can also track the movement of the entire left ventricle wall.  In other words, we want to be able to visualize the left ventricle wall thickness, not just the surface, as this is more suitable for visually detecting infarcted and/or ischemic muscle.

        So, let us describe the notion of tagged MR images a little more in detail. Tagged images appear with a spatially encoded pattern that moves with the tissue. By selectively suppressing the MR signal in planar regions of the heart and then imaging a slice perpendicular to the planes, patterns such as those shown in figure 1 are created. Basically, the patterns or tag lines, which are the intersections between the tag planes and the image slice, appear as black and somewhat parallel lines.  The actual tag lines deform over a short time, usually a cardiac cycle, as the lines are practically straight right after end-diastole and then curve over time.  Since the tag lines are originally straight, the deformed tag lines measure the displacement of the left ventricle. A typical set of images consists of fourteen short-axis image planes, of which seven are imaged with tag planes perpendicular to the x-axis and another seven with tag planes perpendicular to the y-axis. In addition, another six long axis planes are collected from the z-axis. At each plane in the x, y, and z direction, usually about ten or so images are collected within a complete cardiac cycle totaling about two hundred images in all.

MR Tagging

Short axis and long axis image planes

        The first thing in our agenda has been to locate the tagged MR images and extract out the actual image and necessary header information. Not all image sets are the same, as the slices between cardiac cycles may vary and or slice thickness. In addition, to our surprise, not all images are shot in order by axis plane. We are not clear why this is the case, but luckily we discovered this minute nuance.  Once we have finished converting all the images in some kind of format, such as JPEG, we plan on applying a snake algorithm to the images in planes and recording the snake points into several 3-dimensional arrays representing the x, y, and z planes.  The first attempt will involve manually applying numerous snake points to the outer wall of the left ventricle, which is circular in shape. In other words, we will not be taking advantage of the tag lines at all. Assuming the snakes work correctly, we should be able to extract the outer wall points in all three planes in order to create a valid looking 3-dimensional moving model. The next attempt will involve tracking the inner wall. We will then use this information to create another model, which will better convey wall thickness. The final attempt will involve also tracking the individual tag lines to get a more complete wire-frame model, as one can appreciate local contractile performance.  At this point, we have a little application that we wrote which will convert the binary Signa images to PPM images and extract pertinent header information. The next major hurdle will be in importing a bunch of Matlab snake algorithm code that I hacked up last quarter to C++ so that it will work faster. We will also have to modify the code considerably to allow numerous snake lines per image and allow the points to wrap around and connect, as in a doughnut. I envision that getting the snakes to work as planned will be quite a chore and I will have to play with gradients again. For one thing, the actual tag lines will probably throw off the snakes on the part where I just want to follow the actual doughnut shape. Since the tag lines are dark, even taking the gradient of the images will produce high frequency information depicting the tag lines. My guess is that the snake algorithm will want to follow the tag lines, which may force us to use them instead of the outer and inner walls. Finally,  we will need to combine the snake algorithm outcomes to form some kind of dynamic wire model.  The actual division of labor will be determined as we work on the project. Todd and I have worked together in the past as so we usually trade off responsibility here and there depending on who achieves the first milestone.

.