Development and evaluation of ultrasound image tracking technology based on Mask R-CNN applied to respiratory motion compensation system

Background: For respiration induced tumor displacement during a radiation therapy, a common method to prevent the extra radiation is image-guided radiation therapy. Moreover, mask region-based convolutional neural networks (Mask R-CNN) is one of the state-of-the-art (SOTA) object detection frameworks capable of conducting object classification, localization, and pixel-level instance segmentation. Methods: We developed a novel ultrasound image tracking technology based on Mask R-CNN for stable tracking of the detected diaphragm motion and applied to the respiratory motion compensation system (RMCS). For training Mask R-CNN, 1800 ultrasonic images of the human diaphragm are collected. Subsequently, an ultrasonic image tracking algorithm was developed to compute the mean pixel coordinates of the diaphragm detected by Mask R-CNN. These calculated coordinates are then utilized by the RMCS for compensation purposes. The tracking similarity verification experiment of mask ultrasonic imaging tracking algorithm (M-UITA) is performed. Results: The correlation between the input signal and the signal tracked by M-UITA was evaluated during the experiment. The average discrete Fréchet distance was less than 4 mm. Subsequently, a respiratory displacement compensation experiment was conducted. The proposed method was compared to UITA, and the compensation rates of three different respiratory signals were calculated and compared. The experimental results showed that the proposed method achieved a 6.22% improvement in compensation rate compared to UITA. Conclusions: This study introduces a novel method called M-UITA, which offers high tracking precision and excellent stability for monitoring diaphragm movement. Additionally, it eliminates the need for manual parameter adjustments during operation, which is an added advantage.