|
内容記述 |
In radiation therapy, fiducial markers implanted in a pa-tient’s body are tracked using X-ray fluoroscopy to estimatetumor positions. However, flexible markers, such as GoldAnchor® (Naslund Medical AB, Sweden), deform within thebody, making conventional template matching challenging.While deep learning offers a promising solution, the ex-tensive collection and annotation of clinical data requiredfor training pose a significant barrier to adoption. To ad-dress this, we propose a tracking framework that utilizes alightweight Siamese CNN trained exclusively on syntheticfluoroscopy images. Our method generates synthetic datasimulating diverse marker deformations under low-contrastand high-noise conditions, employs dynamic programmingfor stable initial detection, and performs real-time trackingwith a particle filter. In evaluations using clinical data, ourmethod achieves a tracking accuracy of 0.42 ± 0.12 pixelsfor prostate cancer cases and 0.97 ± 0.53 pixels for pan-creatic cancer cases. This significantly outperforms con-ventional methods, particularly in challenging low-contrastpancreatic cancer cases. With TensorRT optimization, theframework achieves a processing speed of 3.8 ms/frame.This work presents a practical solution for high-accuracytracking, reducing data collection costs and facilitating theuse of deep learning in clinical applications. |
