Click here to


Are you sure ?

Yes, do it No, cancel

Probabilistic Decomposition of X-Ray Image Sequence to Extract Obscure Target Objects for Monitoring Intrafractional Organ Motion

M Shindo1*, K Ichiji1 , N Homma2 , X Zhang3 , Y Takai4 , M Yoshizawa5 , (1) Graduate School of Engineering, Tohoku University, Sendai, Japan, (2) Tohoku University Graduate School of Medicine, Sendai, Japan, (3) National Institute of Technology, Sendai College, Sendai, Japan, (4) Southern Tohoku BNCT Research Center, Koriyama, Fukushima, (5) Cyberscience Center, Tohoku University, Sendai, Japan


(Tuesday, 7/31/2018) 7:30 AM - 9:30 AM

Room: Room 205

Purpose: This study is aiming to develop a technique to decompose intensity of a single energy kV X-ray image sequence to extract subset image sequences of moving target tumor, tissues and structures for effective monitoring of internal organ motion during radiation therapy.

Methods: In this study, X-ray image intensity was modeled as a cumulative composite of hidden subset intensities of different objects including moving target tumor, tissues, and rigid structures such as bone. To preliminarily estimate and simulate the relationship between possible combination of hidden subset intensities (PCHSI) and observable intensities of X-ray image sequence, a set of digitally reconstructed radiographs were generated from subset volumes of interest in 4DCT. The PCHSIs and the observable intensities estimated at each pixel were used to compose hidden Markov model (HMM). In HMM, the hidden states, i.e., PCHSI, probabilistically transition among them and probabilistically output the observable intensity. Baum-Welch algorithm was applied to observed intensity sequence to estimate those probabilities. By using Viterbi algorithm, PCHSI sequence were inversely estimated from observed X-ray image intensity sequence at each pixel. The estimated PCHSI sequences were then decomposed into the subset image sequences hidden in the observed X-ray image sequence.

Results: The proposed method was tested on three datasets of non-clinical and two clinical X-ray image sequences. To evaluate the performance of the proposed method, moving target tumor was tracked on both the extracted and original (raw) sequences. The tracking errors on the extracted sequences were 1.71+/-0.52 mm, 2.21+/-1.04 mm, and 1.61+/-0.51 mm, respectively, and the tracking errors on the original sequences were 4.01+/-2.06 mm, 2.39+/-1.12 mm, and 2.61+/-0.43 mm respectively.

Conclusion: We developed an HMM-based target image extraction method from X-ray image sequence. The method can improve the accuracy on target tracking and thus has potential to contribute to effective monitoring of intrafractional organ motion.

Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by Varian Medical Systems, Palo Alto CA and Japan Society for the Promotion of Science (JSPS) Kakenhi.


Image-guided Therapy, Image Processing, Target Localization


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

Contact Email