Atrial fibrillation (AF), which is caused by unsynchronized electrical activity in the atrial chambers of the heart, is a rapidly growing problem in modern societies. Electrical cardioversion and antiarrhythmic drugs can be used to control this condition but it involves side effects and only 40-60% of the AF population is maintained in regular rhythm one year after such treatment. An alternative treatment is catheter ablation in which the sources of electrical dyssynchrony are suppressed by ablating specific tissues using a special catheter. Ablation is performed by exposing the tissue to a high temperature, which results in lesion (scar) formation. This procedure, if successful, provides a final cure to AF.
Magnetic resonance imaging (MRI) has been used for both pre- and and post-ablation assessment of the atrial wall. The MRI assessment can aid in selecting the right candidate for the ablation procedure, and also in the assessment of the post-ablation scar formation. Image processing techniques can be used for automatic segmentation of the atrial wall, which facilitates an accurate statistical assessment of the region. As a first step toward the general solution to the computer-assisted segmentation of the left atrial wall, we propose a new scheme which uses statistical shape learning and segmentation driven by regional statistics to segment the epicardial wall of the left atrium.
Our proposed method is composed of shape registration, shape learning, and image segmentation. Given a training set of binary images corresponding to the segmentations of the epicardial wall of the left atrium, each shape is registered to a fixed arbitrary shape in the training set using a mean-square error registration scheme. Next, we use principal component analysis to learn the shapes. Finally, the shape information available from the learning stage is used to segment the epicardial wall, where the segmentation is driven by local regional statistics. We trained the algorithm on 15 segmentations of the epicardial wall of the left atrium performed by a human expert. The first 6 dominant principal components were used in the statistical shape learning. The algorithm was then used to segment the epicardial wall, as shown in the figure.
Georgia Tech: Behnood Gholami, Yi Gao, Wassim Haddad, and Allen Tannenbaum University of Utah: Rob MacLeod, and Josh Blauer