Taking a multicore DSP approach to medical ultrasound beamforming

Robert Krutsch

July 5, 2011

Robert Krutsch

Implementing the B-MODE
We propose a B-Mode library composed of the following four modules:

1) Envelope detection, decimation and logarithmic compression (ED)
2) Scan Conversion (SC)
3) Median Filtering (MF)
4) Histogram equalization (HQ)

To obtain B-Mode images only the first two modules (ED, SC) are necessary; the next two modules are classic image enhancing techniques that might bring a plus in image quality. Integrating enhancers in the acquisition timeframe has the advantage that no loss in frames/second is observed compared to applying them after the image is already formed. In the following we will speak about bilinear interpolation based Scan Conversion with an output image size of 640x480 and 3x3 window median filtering.

In the literature one can find different approaches to perform envelope detection, based on Hilbert transform, quadrature demodulation or by simply squaring the input signal. We propose an envelope detection algorithm formed from the following sub-modules:

1) Square value of the input beam-formed signal (Figure 12 below)
2) Decimation with a CIC and compensation FIR
3) Logarithmic compression

This method is specific for B-Mode image forming and proves to be one of the friendliest approaches from cycle count consumption perspective. MSC8156 is helping to implement this approach by featuring native adders on 40 bit, which can handle the bit-growth demands of CIC filters.

Figure 12 – In the top figure the spectrum of the echo signal is presented; in the bottom figure the spectrum of the squared echo signal – the band of interest moves towards 0 Hz

In Figure 13 below one can observe the time frame allocated for ED and that a single core is assigned for the processing.

Figure 13 – ED time frame allocation and number of cores used
Median filtering and scan conversion need access to at least three scan lines and respectively two scan lines to be called, this makes them not very suitable for low end FPGAs that have limited buffering resources. In the following, four instances of median filtering and four instances of scan conversion are called on for cores, when enough data is acquired (Figure 14 below).

Figure 14 - MF and SC time frame allocation and number of cores used
Histogram equalization is a slightly different type of algorithm compared to the ones already discussed, due to the fact that at in some point one needs access to the entire image. To maximize the performance, histogram equalization (Figure 15 below) is split into two phases:

(HQ1) Computing the histogram – is executed in the Acquisition and IO timeframe
(HQ2) Apply transformation – is executed when the whole image is acquisitioned (will drop the number of frames/second to about 44.5)


Figure 15 –HQ1 and HQ2 time frame allocation and number of cores used

Bandwidth requirements and cycle count requirements (as obtained after simulator validation) have been summarized in Table 4 below.

Table 4 – Bandwidth and cycle count requirements for B-Mode library
In about 75% of the time frame we will not execute Median Filtering, Scan Conversion and HQ1 so we see a utilization factor as presented in Table 5 below while for the case when MF , SC and HQ1 are called the utilization is presented in Table 6 below.

Table 5.Utilization when MF, SC and HQ1 are not called

Table 6 – Utilization when MF, SC and HQ1 are called
By averaging the utilization from tables… and taking into consideration the unused timeframes a 38% utilization can be observed, about 2300MCPS (million cycles per second) from 6000MCPS available.

Scheduling image scans
Figure 16 below presents the processing performed on Core0 for each scan line; as can be observed 30% of the time frame is unused and cold be utilized for Doppler imaging modes.

Figure 16 – Core 0 utilization
Cores 1 to 4 typically do not execute MF, SC and HQ1 so almost 60% of the time frame is free for other imaging modes (Figure 17 below). When also MF, SC and HQ1 are called, 10% of the time frame is unused (Figure 18 below).

Figure 17 – Core1-4, when MF, SC and HQ1 are not called

Figure 18 – Core1-4, when MF, SC and HQ1 are called
The DDR bandwidth requirements for all the time frames are pictured in Figure 19a below. To lower the cost of the overall system by using a single DDR the unused time frame can be distributed to scan conversion and beamforming Phase 2 , leading to DDR bandwidth requirements acceptable for a single DDR (Figure19b below).

Figure 19 – DDR Bandwidth requirements

Conclusion
In the past three years digital signal processors have progressed significantly in terms of processing power and power consumption making them more and more suitable for low and mid end portable medical ultrasound devices.

The plus in flexibility by using a software approach for all signal processing modules, lower time to market and lower cost by using “of the shelf” hardware that is tested and for that a lot of “getting started” material exist are the big plus of DSPs.

For a simple beam approach, as presented in this paper, a usage fact of about 38% is very encouraging to also move to multiple line acquisition techniques and to add other imaging modes like Power Doppler and Color Doppler.

MAPLE-B, while not used in this exposition, can be found very useful for Spectral Doppler imaging modes and for frequency domain beam forming algorithms.

Robert Krutsch is a DSP Software Engineer at Freescale Semiconductor. He received a BSc degree specializing in automatics and computer science from the Polytechnic University of Timisoara; a BSc degree from the Automation Institute at the University of Bremen and PhD in automatic control from Polytechnic University of Bucharest.

References
[1] Thomas L. Szabo, “Diagnostic Ultrasound Imaging : Inside Out”, Elsevier Academic Press, 2004
[2] Kai E. Thomenius, “Evolution of ultrasound beamformers”, IEEE Ultrasonics Symposium, 1996
[3] Hui Li, Dong C. Liu, “An embedded high performance Ultrasonic Signal processing subsystem”, IEEE computer society, 2009
[4] J. Nelson Wright,” Image formation in diagnostic ultrasound”, IEEE International Ultrasonics Symposium, 1997
[5] Tore Gruner Bjastad, “High frame rate ultrasound imaging using parallel beamforming”, PhD. Thesis, Norwegian University of Science and Technology, 2009
[6]  Iulian Talpiga, “Optimizing Serial Rapid IO Controller Performance in the MSC815x and MSC825x DSPs”, Application Note AN3872, 2010