Semi-Supervised Training of Optical Flow Convolutional Neural Networks\n in Ultrasound Elastography

Convolutional Neural Networks (CNN) have been found to have great potential\nin optical flow problems thanks to an abundance of data available for training\na deep network. The displacement estimation step in UltraSound Elastography\n(USE) can be viewed as an optical flow problem. Despite the high performance of\nCNNs in optical flow, they have been rarely used for USE due to unique\nchallenges that both input and output of USE networks impose. Ultrasound data\nhas much higher high-frequency content compared to natural images. The outputs\nare also drastically different, where displacement values in USE are often\nsmooth without sharp motions or discontinuities. The general trend is currently\nto use pre-trained networks and fine-tune them on a small simulation ultrasound\ndatabase. However, realistic ultrasound simulation is computationally\nexpensive. Also, the simulation techniques do not model complex motions,\nnonlinear and frequency-dependent acoustics, and many sources of artifact in\nultrasound imaging. Herein, we propose an unsupervised fine-tuning technique\nwhich enables us to employ a large unlabeled dataset for fine-tuning of a CNN\noptical flow network. We show that the proposed unsupervised fine-tuning method\nsubstantially improves the performance of the network and reduces the artifacts\ngenerated by networks trained on computer vision databases.\n