Encrypted login | home

Program Information

Deep Boltzmann Machines-Driven Level Set Method for Heart Motion Tracking Using Cine MRI

J Wu

J Wu1*, T Mazur1 , H Gach1 , N Daniel1 , H Lashmett1 , L Ochoa1 , I Zoberi1 , B McClain1 , C Lian2 , S Ruan2 , M Anastasio1 , S Mutic1 , M Thomas1 , H Li1 , (1) Washington University in St. Louis, Saint Louis, MO, USA, (2) University of Rouen, Rouen, Normandy, France.


TU-FG-605-10 (Tuesday, August 1, 2017) 1:45 PM - 3:45 PM Room: 605

Purpose: Dynamic cardiac MRI cine images allow physicians to observe heart motion and provide useful information for analyzing radiation-induced cardiotoxicity and establishing appropriate heart motion management strategies. Nevertheless, high anatomical complexity and relatively poor cine image contrast/resolution have complicated automatic motion evaluation. We propose a deep generative shape model-driven level set method for delineating and analyzing heart motion on cine MRI images.

Methods: The proposed method includes heart shape model training and frame-by-frame heart motion tracking. Considering the geometric shape changes of the heart and the spatial relationship of its neighboring structures, we first use a three-layered Deep Boltzmann Machines (DBM) to train a heart shape model that characterizes both global and local heart shape variations. Then, the shape priors generated by the trained heart shape model are incorporated into a distance-regularized level set-based segmentation method to guide frame-by-frame heart segmentation on cine MRI images. The method was applied to fourteen 1.5 T coronal cine MRI image sequences from seven volunteers and its performance was compared to four other methods.

Results: The tracking accuracy was validated by comparing the results to the average of two manual delineations (used as ground truth). The proposed method achieved an average dice similarity coefficient (89.55% ± 2.26%) for seven breath-hold cine sequences, 88.42% ± 2.49% for seven free-breathing cine sequences, and 88.99% ± 2.38% for all cine sequences. The overall mean margin error was 0.29 ± 0.17 mm for all tested sequences. The comparison with four other methods illustrated the superior tracking performance of the proposed method.

Conclusion: The DBM-trained shape model can characterize both global and local shape properties of the heart for realistic object generation. The proposed method has the potential for assisting in heart motion pattern analysis, and the evaluation of potential heart toxicity induced by external beam radiation treatment.

Contact Email: