Overcoming dropout while segmenting cardiac ultrasound images

Cardiac ultrasound images often contain significant dropout. Image segmentation in the presence of dropout has been previously attempted with a shape prior. However, just by themselves, shape priors fail when the dropout gaps are large. This paper suggests that the dual strategy of modelling the dropout and using shape priors is able to overcome this limitation. In cardiac imaging, dropout tends to occur in predictable regions of the image. The segmentation strategy proposed in this paper learns the dropout function from a training set and uses it as a prior in a maximum a posteriori (MAP) active contour level set formulation. Experimental evidence is provided to show that shape priors can fail in large gaps while the combined strategy is able to overcome this limitation. Comparison of the algorithm segmentation with manual segmentation is also provided