Implementation of Asymmetric Kernel Median Filtering for Real-Time Ultrasound Imaging

Ultrasound images contains speckles that could be considered as salt-and-pepper noise. Median filtering can efficiently remove the noise without blurring the edges. However, median filtering is computationally intensive and not suitable for realtime imaging. In this paper, median filtering on the intermediate ultrasound data using asymmetric median kernel sizes is proposed. To reduce the computational time, the median selection networks with compare-and-swap stages are applied instead of reducing the redundancy of the median calculation among the adjacent image pixels. Moreover, the modification of the compare-and-swap stages to suit for the GPU is also proposed. The computational times using CPU and GPU programs are compared and CNRs of the ultrasound images are also reported.