Clutter Filtering Using a 3D Deep Convolutional Neural Network

Echocardiographic imaging is a well-established clinical modality to assess cardiac morphology and function. However, accuracy of visual reading as well as performance of algorithms designed to measure cardiac characteristics, and as a consequence the diagnostic value of the echocardiographic imaging, can significantly be degraded by the presence of image artifacts such as (stationary) reverberations. Although several filtering techniques have been proposed in the literature to tackle this problem, the current study sought to investigate whether deep learning can offer better artifact removal. Given the spatiotemporal nature of the reverberation artifacts, a 3D (i.e. 2D plus time) deep convolutional neural network (3D U-Net) was trained to filter simulated artifacts that were superimposed onto ultra-realistic synthetic 2D echocardiographic sequences. Performance of 3D U-Net was compared with the singular value decomposition (SVD) technique. The obtained results confirmed the utility of deep learning for artifact filtering and showed its advantage over SVD in handling moving artifacts.