The Application of KAZE Features to the Classification Echocardiogram Videos

In the computer vision field, both approaches of SIFT and SURF are prevalent in the extraction of scale-invariant points and have demonstrated a number of advantages. However, when they are applied to medical images with relevant low contrast between target structures and surrounding regions, these approaches lack the ability to distinguish salient features. Therefore, this research proposes a different approach by extracting feature points using the emerging method of KAZE. As such, to categorise a collection of video images of echocardiograms, KAZE feature points, coupled with three popular representation methods, are addressed in this paper, which includes the bag of words (BOW), sparse coding, and Fisher vector (FV). In comparison with the SIFT features represented using Sparse coding approach that gives 72 % overall performance on the classification of eight viewpoints, KAZE feature integrated with either BOW, sparse coding or FV improves the performance significantly with the accuracy being 81.09 %, 78.85 % and 80.8 % respectively. When it comes to distinguish only three primary view locations, 97.44 % accuracy can be achieved when employing the approach of KAZE whereas 90 % accuracy is realised while applying SIFT features.