Stochastic Systems Group  

Walter Sun  SSG, MIT
Segmenting the left ventricle and tracking it across a cardiac cycle can help cardiologists diagnose the health of a heart. Because the manual segmentation of a dynamic sequence is extremely timeconsuming, accurate automatic tracking procedures are desired. However, a major challenge to tracking the left ventricle is the fact that the chambers in the heart deform in a complex manner, making simple dynamic models insufficient for characterizing their evolution. In this talk, we propose a method to automatically track the left ventricle. We first introduce a probabilistic framework which allows us to recursively estimate the state (the left ventricle boundaries) of the system at each frame. Within this formulation, the dynamics of the system are needed. Since such system dynamics are often unknown, we discuss an information theoretic approach which learns the dynamics. In particular, we choose the parameters to a function f which maximizes the mutual information between X_{t} and f(X_{t1}).
We present results using a curve evolution approximation to the formulation that yields a single estimate at each frame. In addition, we present a more general implementation using particle filters to approximate the recursive estimation process. Using this method, we obtain samples which represent the distribution of curves at each frame.
Problems with this site should be emailed to jonesb@mit.edu