Context

We chose to tackle biopsy needle steering. Bevel tip needles move in the direction normal to the bevel when inserted, allowing physicians to navigate hard to reach targets in the body. They can change the direction of the needle to avoid obstacles by rotating the needle base (and thus the bevel tip), causing the body of the flexible needle to deflect.

We modeled the problem as a partially observable Markov decision process (POMDP) to generate an optimal policy for needle orientation at each step of insertion to reach the target while cutting as little tissue as possible. For the discrete POMDP, the transition model allowed for 3 possible state transitions at each step, and an uncertainty of 9 possible tip locations. For the continuous POMDP (solved by POMCP), we used Gaussian distributions that an ultrasound might observe, and beam-bending mechanics for transition states.

Report (pdf)

Skills: Reinforcement learning, Julia language, Git

Needle Steering