Multi-processor system for real-time deconvolution and flow estimation in medical ultrasound

More and more advanced algorithms are being introduced for performing signal and image processing on medical ultrasound signals. The algorithms often use the RF ultrasound signal and perform adaptive signal processing. Two examples are the cross-correlation estimator for blood velocity estimation and adaptive blind deconvolution. The first algorithm uses the RF signal from a number of pulse emissions and correlates segments within different pulse-echo lines to obtain a velocity estimate. Real-time processing makes it necessary to perform around 600 million multiplications and additions per second for this algorithm. This has until now only been possible by using the sign of the signals, and such an implementation does not give optimal performance. The second algorithm also uses the RF data, and first performs an estimation of the one-dimensional pulse in the tissue as a function of depth. Then a Kalman filter is used with a second time-reversed recursive estimation step. Here it is necessary to perform about 70 arithmetic operations per RF sample or about 1 billion operations per second for real-time deconvolution. Furthermore, these have to be floating point operations due to the adaptive nature of the algorithms. Many of the algorithms can only be properly evaluated in a clinical setting with real-time processing, which generally cannot be done with conventional equipment. This paper therefore presents a multi-processor system capable of performing 1.2 billion floating point operations per second on RF ultrasound signals. It consists of 16 ADSP 21060 processors each capable of 80 Mflops. Four processors are placed on one board with 24 MBytes external storage and an internal storage of 0.5 MBytes per processor. All processors can access all storage on its physical board, and are further connected through parallel interface channels. Each channel can transmit 40 MBytes a second without slowing the processor down, and each processor has 6 of these channels. Four of these are accessible through front panel connectors, so that an almost arbitrary network of the 16 processors can be made. The system has been interfaced to our previously-developed real-time sampling system that can acquire RF data at a rate of 20 MHz and simultaneously transmit the data at 20 MHz to the processing system via several parallel channels. These two systems can, thus, perform real-time processing of ultrasound data. The advantage of the system is its generous input/output bandwidth,