CNN‐based spatio‐temporal nonlinear filtering and endocardial boundary detection in echocardiography

In this paper, a CNN based spatio-temporal approach is introduced for finding the endocardial (inner) boundary of the left ventricle from a sequence of echocardiographic images. The discussed analogic1 CNN algorithm combines optimal nonlinear filtering and constrained wave propagation in order to estimate the continuous contour of a moving object in a medium where the edges are ill-defined. In the preprocessing phase, nonlinear filtering is employed to remove the coherent speckle noise that corrupts the images. It is verified that an optimal filtering strategy should estimate the mode of the local intensity histogram. Three different approximations of the mode filter were implemented, derived from robust statistics2 and geometry-driven diffusion3, that give an output consistent with the maximum likelihood estimate of the noisy sequence. The kernel of the left ventricle is located and the boundary is found using a fuzzy-adaptive technique that embodies constrained wave propagation. Boundary dislocation, area and smoothness constraints are transformed into the transient length of the CNN while the a priori knowledge about the heart morphology is built into the spatial template parameters (weight values). Special emphasis is given to VLSI implementation complexity. It is shown that the core of the algorithm can be realized using the already available CNN chips. Furthermore, it is argued that all templates of the complete solution belong to the implementation frame that is considered for the next generation of CNN Universal Chips. This study demonstrates that the discussed novel approach allows a reliable endocardial boundary tracking of the left ventricle in real-time using a spatio-temporal CNN visual microprocessor.