Deep Imitation Learning for Automated Drop-In Gamma Probe Manipulation
Kaizhong Deng, Baoru Huang, Daniel S. Elson · 2023
Prostate cancer is one of the most common cancers in the UK, and Robotic-Assisted Surgery (RAS) has become a common method for prostate cancer surgery. Sentinel lymph node biopsy (SLNB) is an important component of prostate cancer surgery and provides accurate diag- nostic evidence of disease extent. A drop-in gamma probe, SENSEI, has been designed to improve the accuracy of sentinel lymph node detection in RAS. An example of its in vivo usage can be seen in Figure 1. It can distinguish cancerous tissue from normal tissue by detecting the radiation emitted from radiolabeled probes that have been injected into the body. A feasibility study has demonstrated that the drop-in gamma probe can provide accurate identifica- tion of positive nodes following the administration of technetium-99m nanocolloid [1]. However, relying on the live gamma level display and audible feedback from the console while the probe is scanned across the tissue surface is not an easy or intuitive way to identify hidden affected lymph nodes. This might affect the effectiveness of less experienced surgeons and latent hot spots may be overlooked. To address these issues, we propose a robotic scanning method to automatically and systematically examine an entire target area and locate the hot spots. In this study, we present a deep imitation training workflow based on simulation data for an end-to-end learning- based agent capable of systematically scanning target areas using visual input and the current robot state. The evaluation result shows that this approach is promising to automatically control the drop-in gamma probe.