A typical cineMRI sequence (obtained from www.isi.uu.nl)
The above is a "short axis" view of the heart, a perspective that clearly shows the right and left ventricles. Given the movie-like qualities of the image sequence, it is then possible to evaluate the volume of the ventricles through an entire cardiac cycle and from this, we can obtain mountains of useful measures. End Diastolic Volume (EDV), End Systolic Volume (ESV), Stroke Volume (SV), Ejection Fraction (EF), Cardiac Output (CO), the wealth of information is endless. But first....how?
"Sigh," says Dr. Weinsaft.
This is done by meticulously tracing each of the frames manually, separating structures of interest by dropping hundreds of points via a steady hand and a trusty mouse. Such an exercise takes an experienced cardiologist over ten minutes. For me? Coupled with a finicky software package that heartlessly punishes mistakes by totally erasing existing work? Better section off an afternoon.
A single "segmented" frame (obtained from www.isi.uu.nl)
Upon repeating the process 30 or more times for each short axis slice, we obtain something like this:
A fully segmented image sequence (obtained from www.isi.uu.nl)
From here, the software package takes over, automatically calculating the relevant statistics from the segments. After copying down this information, Dr. Weinsaft begins his clinical reading.
The problem? Segmenting the heart manually is a time-consuming, monotonous affair that can often take more time than the reading itself. If only there was an automated method...
Fortunately, there is. An algorithm, developed by Noel Codella, of Cornell University, is capable of segmenting the inner region of the left ventricle, with future development aimed towards fully segmenting all regions of interest in the heart. In addition to being far more consistent in segmentation decisions (an admittedly subjective judgement call at times when done manually), Noel's algorithm is fast. The time required for evaluating a single case is 1/20th that of a manual evaluation by an experienced observer.
Unfortunately, the algorithm is still being fully evaluated and its results cannot be fully relied on as of now. My job, for the remainder of the summer, is to segment as many cases as possible, using both the automatic algorithm and manual segmentation. By quantifying the agreement between the two methods and by measuring useful statistics such as the amount of time saved using automation, it is hoped that we'll provide compelling evidence as to the accuracy and speed of computer-assisted evaluations.
This convergence between computerized automation and "traditional" human evaluation is one of the aspects of engineering that interest me the most. A cardiologist shouldn't have to spend more time connecting-the-dots than than he does in rendering a diagnosis. An experienced clinician with decades of schooling and experienced shouldn't be limited by how quickly they can perform an activity we learned in pre-school. And with a little more work from engineers, hopefully they won't have to.