GPU accelerated implementation of ultrasound radio-frequency time series analysis

Abstract

The ultrasound radio-frequency (RF) time series method has been shown to be an effective approach for accurate tissue classification and cancer detection. Previous studies of the RF time series method were based on a serial MATLAB implementation of feature calculation that involved long running times. Clinical applications of the RF time series method require a fast and efficient implementation that enables realistic imaging studies within a short time frame. In this paper, a parallel implementation of the RF time series method is developed to support clinical ultrasound imaging studies. The parallel implementation uses a Graphics Processing Unit (GPU) to compute the tissue classification features of the RF time series method. Moreover, efficient graphical representations of the RF times series features are obtained using the Qt framework. Tread computing is used to concurrently compute and visualize the RF time series features. The parallel implementation of the RF time series is evaluated for various configurations of number of frames and number of scan lines per frame acquired in an imaging study. Results demonstrate that the parallel implementation enables imaging of tissue classification at interactive time. A typical RF time series of 128 frames and 128 scan lines per frame, the parallel implementation be processed in 0.8128 ± 0.0420 sec.