Cardiovascular diseases: using AI to navigate the catheter

Treatment of a heart attack or stroke caused by vascular occlusion must be prompt to prevent further damage to vital tissue. Unfortunately, the actual treatment is often preceded by a lengthy catheter-based procedure where the cardiologist manually guides the catheter to the affected vessel. AI might perform this task in the future.

In this MEDICA-tradefair.com interview, Lennart Karstensen talks about the development of AI-powered catheters, explains how the algorithm learns to navigate inside the vessels and reveals where the application could be used in the future.

Mr. Karstensen, you are working on setting up cardiac catheters with AI-powered navigation. How did you come up with the idea?

Lennart Karstensen: Catheter navigation is a complicated procedure, especially in the case of strokes, but also heart attacks. In difficult cases, the surgeon takes up to an hour to guide the catheter to the right site. Our project was created to assist and relieve the surgeon with the help of autonomous catheter navigation. The idea here is that the physician would then merely have to supervise the navigation process versus actively performing the task. This would allow him or her to focus more on the diagnosis and the actual treatment, which follows this process.

How does the navigation process work?

Karstensen: The artificial neural network learns how to use a manipulator to twist and slide the guidewire to steer it to its target destination. Here’s what this could look like in the future: once the catheter has been placed at the start of the procedure, X-rays are taken. A CT scan that was taken before the procedure also illustrates the patient’s vascular tree. Our project partner, the Fraunhofer Institute for Digital Medicine MEVIS, has developed software for catheter tracking that tracks the position of the catheter inside the body.

Armed with these three pieces of information – the position of the catheter, the illustration of the vascular tree and the aim of the intervention – the goal is to guide the catheter autonomously to the target site. At present, this is still happening via phantoms, but the process already works reliably in this setting.

Does this mean you no longer require imaging during the procedure?

Karstensen: Not quite since you have to continually check the position of the catheter. Having said that, this process eliminates many doses of contrast medium administration, which the surgeon needs to visually track the catheter and find branch vessels. The neural network is able to remember the current position and the vessels with the accuracy of a computer and can proceed to navigate from there.

How do you train the AI model?

Karstensen: At the moment, this is done purely using a software model that simulates the catheter navigation through the vascular tree. In this instance, we choose a target and the algorithm tries to reach it. If it can correctly do this, it is awarded a plus point. If it is incorrect, a minus point is awarded. Initially, random actions lead to the target. The neural network remembers the correct moves and continually adapts. Gradually, it becomes more efficient as it reaches the target more often and repeats the correct action at the right moment.

This closely resembles the way someone learns a new sport: trial and error eventually leads to success and is replaced by targeted and purposeful actions. That is also why we rely on simulation to test actions frequently, in parallel and promptly. Phantoms can be too expensive and time-consuming.

If the navigation learns through simulation, how can it take differences in patient anatomy into account?

Karstensen: The simulation currently has only one vascular tree geometry. We are working on training the algorithm to navigate through several types of geometry. Once we have trained the neural network on 30-40 geometries of the vascular tree, it should be able to master other vascular trees or learn very quickly how to navigate them. After all, CT scan segmentation actually provides information about the specific patient. The vascular tree only has certain variations, which means there is not an endless array of differences when it comes to patients. Besides, the physician would intervene in special cases anyway. Yet mostly it comes down to whether branches are located somewhat to the right or more to the left.

Eventually, it should learn to navigate any type of vascular tree based on the geometry information we provide. This is the focus of our research in the next six months. We may be able to present something to this effect at the MEDICA 2020 trade fair.

What are the next steps in your project?

Karstensen: We are currently applying for project funding from the German Federal Ministry of Health to advance the neural network for cardiovascular applications. This means that it must be made compatible with procedures performed in a cardiac catheterization lab. We would then have proof of concept, demonstrating feasibility in the cardiovascular setting based on the available information and generated images - with the caveat that we only tested it with phantoms and in simulation. The next step would be to create a finished product as part of an industry partnership or a spin-off company.