Real-time processing in dynamic ultrasound elastography: A GPU-based implementation using CUDA

This paper addresses the computational cost of the normalized cross-correlation (NCC) algorithm in ultrasound elastography. Parallel implementations of the NCC algorithm based on multicore architectures and a graphical processor unit (GPU) are formulated and applied to radio-frequency (RF) data from dynamic elastography experiments. Compared to single computer processor unit (CPU) performances, results show that parallel implementation of the NCC algorithm allows speedups of less than 5 for multi-threaded execution on CPU and up to 85 using a GPU. Processing frame rates from 80 to 173 sec^-1 have been achieved for large fields of view with good spatial resolution. The trade-off between accuracy, spatial resolution and computational cost in displacement estimation using the NCC algorithm therefore appears obsolete. Open source codes for implementing the NCC algorithm on GPU are made available at www.lbum-crchum.com.