Learning-based Target Localization in Robotic Radioguided Surgery in Noisy Environments

Abstract In radioguided surgery, G-probes are used intraoperatively to localize targets marked by radionuclides. However, the interpretation of G-probe measurements is challenging due to background activity from surrounding organs. This work investigated whether deep neural networks can localize radiation sources in such environments and how different background activities impact this task. A physics-guided forward model simulated G-probe measurements for different intra-abdominal distributions, including an anatomically inspired bladder activity. A convolutional neural network was trained on simulated measurements to predict 3D target source positions. Results indicated that prediction accuracy improved with more Gprobe measurements and degraded with increased background activity. In particular, proximity to high-activity regions like the bladder significantly reduced accuracy. This study demonstrates the need to consider background activity distributions for target localization and that a convolutional neural network could solve this task.