Abstract

Info

Title:: SAM2-based automatic tumor segmentation and real-time tracking in dynamic imaging for lung cancer radiotherapy

Author(s):: ZHU Yongqi; ZHANG Hua; School of Biomedical Engineering, Southern Medical University, Guangzhou 510515, China

Keywords:: Keywords: lung cancer radiotherapy Segment Anything Model 2 respiratory motion tumor tracking

Abstract:: Abstract: To address the challenge in lung cancer radiotherapy where complex thoracic anatomical backgrounds lead to low tumor contrast in X-ray images and limited tracking accuracy, a Segment Anything Model 2 (SAM2) is introduced into lung tumor tracking for radiotherapy. Leveraging the core advantage of prompt-driven visual tracking capability, a tumor recognition and segmentation strategy suitable for clinical scenarios is developed, enabling consistent identification and continuous tracking of tumor regions in both static and dynamic images. The method is evaluated using 3 types of data, namely simulation data, phantom data, and clinical data. Results show that on simulation data, the model achieves an average inference speed of 11.5 FPS, and a mean Dice coefficient of 0.730 7, while maintaining favorable segmentation consistency across frames. In phantom experiments, 10 prompt clicks (5 foreground points and 5 background points) are found to provide the optimal segmentation setting. On clinical data, the model produces stable segmentation results that closely match the actual tumor morphology. The study indicates that SAM2 can address key difficulties in tracking tumor motion during lung cancer radiotherapy, providing technical support for precise lung cancer radiotherapy.