Premium
WE‐A‐134‐03: A Kernel‐Based Method for Non‐Rigid Tumor Tracking in KV Image Sequence
Author(s) -
Zhang X,
Homma N,
Ichiji K,
Takai Y,
Narita Y,
Abe M,
Sugita N,
Yoshizawa M
Publication year - 2013
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.4815509
Subject(s) - bhattacharyya distance , artificial intelligence , computer vision , image guided radiation therapy , kernel (algebra) , histogram , computer science , histogram equalization , feature (linguistics) , histogram matching , frame (networking) , pattern recognition (psychology) , mathematics , algorithm , medical imaging , image (mathematics) , telecommunications , linguistics , philosophy , combinatorics
Purpose: To develop a fast algorithm to track the non‐rigid lung tumor motion in KV X‐ray image sequence for image‐guided radiation therapy (IGRT). Methods: The KV X‐ray image sequence was acquired on the Varian On‐Board Imager (OBI) KV imaging system. As a pre‐processing, a histogram equalization was employed to enhance the tumor contrast in the images. In the first frame, a target model containing tumor area was delineated manually, and its feature space was represented by its histogram weighted with an isotropic kernel. In the subsequent frames, the tumor location was estimated by maximizing a Bhattacharyya coefficient which measures the similarity between the target candidates in the current frame and the target model in the previous frame. The numerical solution of maximizing the Bhattacharyya coefficient was performed by using a mean‐shift algorithm. Results: We implemented four conventional template matching algorithms to compare their performance with the proposed method. Experiments were conducted on four lung tumor kV image sequences of resolution 0.26 mm/pixel. Each sequence consists of 100 frames. The ground truths of the tumor motion were obtained by manual localization. Experimental results demonstrated that the proposed algorithm was superior to the conventional template matching algorithms in terms of its accuracy and computational cost. Conclusion: This study aims at developing a robust and fast algorithm used for tracking the lung tumor for real‐time IGRT. Due to the histogram representation of the target feature, the proposed method is robust against the tumor's shape deformation. In addition, the proposed tracking algorithm is based on a kernel gradient estimation and its computational cost is much lower than that of the conventional template matching algorithms that involve in exhaustive search procedures. The proposed method shows the effectiveness of tracking tumor in KV image sequence and a promising prospect for MV image sequence.