Enhanced motion estimation by training a deep learning optical flow algorithm on a hybrid dataset

Cardiovascular diseases (CVDs) are the primary cause of death worldwide. Cardiac ultrasound (US) is widely used to assess CVDs, allowing to evaluate regional myocardial function through the quantification of regional motion and deformation. Speckle tracking is the most widely accepted method for cardiac motion estimation (ME). However, these methods face challenges due to ultrasound limitations, such as speckle decorrelation. This work proposes a deep learning (DL) ME solution based on the PWC-Net architecture. To improve ME robustness, we propose to augment its training with synthetic 2D B-mode sequences generated using a fast convolution-based ultrasound simulator (the COLE simulator). Hence, two datasets were employed to train PWC-Net, one synthetic, and one In-vivo, with 100 and 116 US recordings respectively, each with their corresponding reference motion used as ground truth. Overall, training with a mixed dataset outperformed a single dataset training regime (pixel-wise end-point error of 0.50 compared to 0.53 and 1.30 when using in-vivo or synthetic US data only), demonstrating the relevance of synthetic data for developing DL-based ME solutions for cardiac US.