Handcrafted features vs ConvNets in 2D echocardiographic images

In this paper, we address the problem of automated pose clas-sification and segmentation of the left ventricle (LV) in 2Dechocardiographic images. For this purpose, we compare twocomplementary approaches. The first one is based on engi-neering ad-hoc features according to the traditional machinelearning paradigm. Namely, we extract phase features to buildan unsupervised LV pose estimator, as well as a global im-age descriptor for view type classification. We also apply theSupervised Descent Method (SDM) to iteratively refine theLV contour. The second approach follows the deep learn-ing framework, where a Convolutional Network (ConvNet)learns the visual features automatically. Our experiments ona large database of apical sequences show that the two ap-proaches yield comparable results on view classification, butSDM outperforms ConvNet on LV segmentation at a signifi-cantly lower training computational cost.