MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Three-Dimensional Tumor Position Estimation Using Image Features On Orthogonal KV X-Ray Projections: Comparison of Two Feature Extraction Algorithms

S Anbo1*, M Nakamura1 , N Mukumoto1 , T Ono1 , T Mizowaki1 , (1) Kyoto University

Presentations

(Tuesday, 7/16/2019) 9:00 AM - 10:00 AM

Room: Pins Room | Hall 2

Purpose: To identify target positions on kV X-ray projections without using internal markers by two feature detection algorithms, such as SURF and KAZE.

Methods: The following experiments were performed with the Vero4DRT which enables orthogonal kV X-ray imaging. In this study, the anthropomorphic chest phantom, which emulated a human torso including the lungs, skeleton and a lung tumor, were used. While the tumor moved with amplitude of 10 mm and breathing cycle of 6 s in cranial-caudal direction, sequential orthogonal kV X-ray projections were acquired. Image acquisition angle pairs (Imager 01, Imager 02) were (0°, 270°) to (90°, 0°) by 10 degree-interval and covering at least one breathing cycle. Enhancement of image contrast and the image feature point detection by KAZE and SURF algorithms were demonstrated on MATLAB. Initially, centroids of the tumor positions were detected manually as reference 2D tumor positions. On KAZE and SURF, a tumor position on i th frame were calculated automatically from the tumor position on i-1 th frame and positional transitions of the feature points, while initial tumor positions were set to the first 2D reference tumor positions. Next, the calculated and reference tumor positions were converted into 3D positions from 2D localization data using predefined calibration parameters. Finally, root-mean-square error (RMSE) between 3D tumor and reference positions were obtained.

Results: Median RMSEs for SURF and KAZE were 4.1 mm (range; 2.0-7.3 mm) and 2.3 mm (range; 1.3-3.1 mm), respectively, except for the angles where RMSEs for several imaging pairs were not obtained due to insufficient imaging contrast to extract effective feature points.

Conclusion: KAZE had better accuracy than SURF to estimate tumor position. More enhancement of imaging contrast would be necessary to improve tracking accuracy.

Keywords

X Rays, Fluoroscopy, Localization

Taxonomy

IM/TH- RT X-ray Imaging: General (most aspects)

Contact Email