Automatic real-time view detection

Abstract

This work presents an algorithm capable of classifying an echocardiographic view as either an apical two chamber view, four chamber view or long axis view. It also provides a score on the overall image quality. The algorithm is based on a deformable non uniform rational B-spline (NURBS) model updated in an extended Kalman filter framework. Models are constructed for each of the three standard views. Each model is updated using a combination of edge and speckle-tracking measurements, where weak edges and edges strongly deviating from their neighbor edges are discarded. The most probable standard view is found using feature detection and general successfulness in detecting edges. This is also used as a measure of overall view quality. The algorithm was trained and validated using 68 recordings from the Norwegian HUNT database. An echocardiographer classified each recording as one of three standard views. 33 randomly chosen recordings, with approximately 10 of each view, were used for training. The other 35 recordings were used for validation. The algorithm successfully classified the view in 32 of 37 cases (86.5%). Each classification is accompanied by a score, which can be used to assess image quality.