Lesion tracking method using CNN for non-invasive ultrasound theranostic system

Our laboratory has developed a system to treat kidney, liver, and other tumors and also stones produced by these organs with high-intensity focused ultrasound while tracking lesions that move due to respiration, heart contractions, and body movement. This system estimates the movement of organs by analyzing the ultrasonic image obtained from the probe and compensates for this motion with robotic controls. In recent years, various medical imaging analysis techniques using deep learning have been proposed and examined, but they have not been sufficiently examined as a technique for tracking a specific organ with a robot system. The purpose of this study was to investigate and verify the performance of Faster R-CNN, which is commonly used for general image recognition, for detecting kidneys in ultrasound images. RegNet, a newly proposed neural network that simultaneously predicts the presence and location of kidneys, has shown superior results in detecting and tracking kidneys in comparison to Faster R-CNN. However, further improvement is necessary because the existence prediction and position detection functions produce independent results.