Catheter's in X-ray

Today I am going to talk about the research I have done in determining catheter tip locations from chest radiographs and a little of the future work we have planned.

As I mentioned previously, a lot of time is spent by radiologists doing intensive care unit imaging to make sure that every tube, line, and catheter is where it is supposed be. This is because if they aren't, complications and even death can occur. For example, its very important that feeding tubes end up in the stomach and not down one of the bronchi before you try to feed someone. But these reads are very time consuming, both in per scan rate and number of scans to read. So my research is focused on developing a method of computer aided detection of the tips, so that we at the very least can reduce the workload of the radiologists, which would further result in an increase in efficiency and throughput.

The first thing we looked at was that catheters were synthetic objects. They are basically tube with the ends cut off (kind of like a straw). This means that
1) The profile is consistent along its length
2) The profile of the object can known a priori
3) Intensity variations can be explained by
But we needed to verify the assumption so we took sample profiles along the length and got the graph below.



The profiles had roughly the same added attenuation at each point, the baseline is what mainly changed. So we are good to go on that front.

Taking into account the synthetic nature of the catheters, we developed a way to generate profiles automatically. These profiles are then matched from points outside the body to the tip, until no more evidence is found. We used normalized cross-correlation because it allowed for matching to occur regardless of intensity variations. Doing this progressive matching gave us results like the image below.



Not bad for proof of concept, right?

Luckily, I got to see many of the steps in how the images I am working with were acquired. This is actually quite helpful in knowing all the potential problems that the algorithm might need to deal with. I also spent time gathering data so that I have a larger set to work with. We are also planning on generating synthetic dataset next week to try out other methods of more direct detection.