Motion Correction Using Deep Learning Neural Networks - Effects of Data Representation

An in-silico investigation of the effects of ultrasound data representation on the accuracy of the motion prediction made using deep learning neural networks was carried out. The representations studied include: linear (‘envelope’), log compressed, linear with phase and log compressed with phase. A UNet model was trained to predict non-rigid deformation field using a fixed and a moving image pair as the input. The results illustrate that the choice of the representation plays an important role on the accuracy of motion estimation. Specifically, representations with phase information outperform the representations without phase. Furthermore, log-compressed data yielded predictions with higher accuracy than the linear data.