Optimal Path Planning of a Multi-tube Bent Cannula for Deep Brain Surgery

By Lina Sorg

A cannula is a thin, hollow tube that is inserted into the body or brain to administer medication, drain fluid, take samples, or introduce a surgical instrument. It typically contains a trocar needle, which pierces the body to reach the targeted space. One valuable application area for cannulae is minimally invasive stereotactic neurosurgery — the process of locating surgical targets in the brain in the context of an external frame of reference. In the case of a tumor, for example, surgeons must access a particular target in order to take a sample. The current state of the art for deep brain stimulation or biopsy involves a straight cannula that enters the skull through a small hole to reach the target position (see Figure 1). During this process, the cannula must avoid contact with obstacles like ventricles and blood vessels. Because cannulae are rigid, invasive open surgery is necessary if surgeons cannot reach the target via a straight path. “This is very detrimental for the patient, especially because the surgeon then has to use his hands to push away the brain,” Matthias Hoffmann of Saarland University said.

During a minisymposium presentation at the 2023 SIAM Conference on Computational Science and Engineering, which took place this week in Amsterdam, the Netherlands, Hoffmann employed sensitivity- and data-based iterative learning control (ILC) to explore the possible use of multi-tube bent cannulae for deep brain neurosurgery. These bent cannulae navigate the brain via a curved path when no straight path is available, thus avoiding open surgery (see Figure 2). Hoffmann utilized concentric tube continuum robots—a series of multiple nested interacting curved tubes—to build up a curved cannula with a common backbone shape. Because each tube has two degrees of freedom, they can be pushed out and rotated as necessary to reach the target. Such cannulae are only viable if they have a target point accuracy of less than one millimeter (mm).

Hoffmann relied on the Rucker model for optimal path planning. This geometrically exact, nonlinear, quasi-static model computes an approximation of a cannula’s backbone based on given tubes and their \(\alpha\) and \(\beta\) values. Neglected effects like contact forces, friction, and material nonlinearities mean that some errors will be unavoidable; nevertheless, the Rucker model serves as a good starting point. “We want to compute an optimal solution for \(\alpha\) and \(\beta\) and adapt it to reach the target point by modifying \(\alpha\) and \(\beta\) based on measurements of the system,” Hoffmann said.

A photogrammetry system allowed Hoffmann to triangulate the position of the cannula tip; this process occurs outside of the patient because the system is robust to changes (see Figure 3). He conducted optimization-based trajectory planning to obtain the \(\alpha\) and \(\beta\) values, measured the position of the cannula tip, and computed the tip position’s sensitivity for \(\alpha\) and \(\beta\) via the Rucker model. Hoffmann then used sensitivity-based ILC to change \(\alpha\) and \(\beta\) and convert to the actual target. “We want the tip to end up in the target at the end,” he said. He was able to change the \(\alpha\) and \(\beta\) levels again and again based on the measurements. The photogrammetry system confirmed that the solution converges.

In the difficult case of especially bent cannula tubes, the algorithm reduced the target error from 37 mm to less than 0.2 mm, which meets the aforementioned surgical requirements. “Sensitivity information for ILC, while a rough approximation, is a valuable tool,” Hoffmann said. The solution reached a minimum in 25 iteration steps and began to deteriorate, but Hoffmann stored it and can return to it once performance worsens. If the curvature is not strong enough, the algorithm reduces the error over time but never reaches the target point. In this case, one must replace the tubes and start again. 

Next, Hoffmann briefly overviewed results from data-based ILC, for which he trained neural networks (NNs) using PyTorch. Results from this method were not as good as those from the sensitivity-based approach; the error reduced from 32 mm to six mm in nine iterations, which does not meet the necessary requirements for surgery. Future steps for this project involve working to improve the data-based outcomes. Hoffmann and his team also intend to employ higher-order sensitives, combine sensitivity and data-based approaches, address constraints, and explore NN-based ILC techniques as a possible initialization for reinforcement learning.