MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Novel Real-Time Markerless Target Tracking Pipeline Based On Faster R-CNN for Lung Cancer Radiotherapy

L Deng1,4*, Z Dai2, X Liang3,4, H Zhao4, H Quan1, Y Xie4, (1) Wuhan University, Wuhan, Hubei, CN, (2) The Second Affiliated Hospital of Guangzhou University of Chinese Medicine, Guangzhou, Guangdong, CN, (3) Stanford Univ School of Medicine, Stanford, CA, (4) Shenzhen Institute Of Advanced Technology,Shenzhen, Guangdong, CN

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Traditional templates matching methods have achieved good accuracy in markerless target tracking on a Linac with the on-board imaging system. Nevertheless, its computational complexity hinders us from marching towards real-time tracking. Therefore, we proposed a novel real-time pipeline with the core module Faster R-CNN to speed up the target locating process on X-ray images. Additional modules of Unscented Kalman Filter and Deep Bi-LSTM followed were intended for estimating target positions and predict respiratory motion.

Methods: To characterize the proposed target locating process, we used a dynamic thorax phantom (CIRS) that simulates the human chest anatomy and target motion in the lung. Tumor surrogates of three clinically relevant sizes (diameter of 12, 18, 27mm), were chosen as test target samples. The targets were programmed to move simply in the inferior-superior direction. The locating error was calculated as the root mean square error (RMSE) between the estimated position and the programmed. To verify the proposed respiratory motion prediction process, we processed a patient-specific respiratory curve at the supine gesture to drive the target's motion. And the forecast error was defined as RMSE between the current programmed position and position predicted from streaming estimated target position results.

Results: For all the success tracking cases, locating errors for three sizes of tumor targets were 0.92mm, 0.85mm, and 0.43mm for 12mm, 18mm, 27mm, respectively. Overall end-to-end processing time of a single fluoroscopic image for locating the target is nearly the same, around 220ms(4.5FPS), for all the target sizes. For a latency of 200ms and 400ms, forecast errors were 0.034mm and 0.09mm.

Conclusion: Our proposed Faster-RCNN based pipeline was qualified for real-time in the case of phantom study for its balance of accuracy and speed. However, more verifications are under circumstances when more complex target motion patterns and anthropomorphically realistic anatomy are considered.

Keywords

Target Localization, Fluoroscopy, Lung

Taxonomy

IM- X-Ray: Machine learning, computer vision

Contact Email