Class-specific regression random forest for accurate extraction of standard planes from 3D echocardiography

This paper proposes a class-specific regression random forest as a fully automatic algorithm for extraction of the standard view planes from 3D echocardiography. We present a natural, continuous parameterization of the plane detection task and address it by the regression voting algorithm. We integrate the voxel class label information into the training of the regression forest to exclude irrelevant classes from voting. This yields a class-specific regression forest. Two objective functions are employed to optimize for both the class label and the class-conditional regression parameters. During testing, high uncertainty class-specific predictors are excluded from voting, maximizing the confidence of the continuous output predictions. The method is validated on a dataset of 25 3D echocardiographic images. Compared to the classic regression forest [1], the class-specific regression forest demonstrates a significant improvement in the accuracy of the detected planes. © 2014 Springer International Publishing Switzerland.