Nội dung

1 Hardware-Software Partitioning of Digital Signal Processing in Ultrasound Medical Devices a Case Study Marcin Lewandowski Institute of Fundamental Technological Research, Polish Academy of Sciences Pawinskiego 5B, 02-106 Warsaw, Poland 1. Introduction The development of ultrasound devices and diagnostic methods is closely related with the development of microelectronics and digital signal processing. In most of the state of the art electronic devices, the signal is digitally processed. These devices may be generally categorized based on the number of parallel processing channels. Imaging devices with multi-element linear or phased arrays usually have 16-256 transmission/reception channels. Single channel processing is usually performed in imaging devices with mechanically moved single element sector heads, the so-called “wobbler” and in dedicated Doppler devices. The methods of signal processing are much the same in all device categories. The paper presents a general overview of ultrasound signal processing and its digital implementation with emphasis on hardware-software partitioning. The available state of the art methods and systems of digital signal processing using both hardware and software are presented as well as the issues pertaining to algorithm implementation methodology. The state of the art system solutions are presented based on the descriptions of representative ultrasound devices, found in literature. The similarities between the presented devices and radio signal processing systems used in telecommunication are also discussed in the paper. Based on device description, the authors present the architecture of processing and communication as well as specific design solutions. The discussed issues and system solutions are analyzed based on two ultrasound medical devices, namely: • uScan - high frequency ultrasonograph with coded transmission (Lewandowski & Nowicki, 2008), • digiTDS - transcranial Doppler system (Lewandowski et al., 2009). Both devices have been designed and built by the author and co-workers at the Institute of Fundamental Technological Research, Polish Academy of Sciences. They have been designed for commercial purposes, in conformity with medical standards and economic limitations. The analysis of the presented solution comprises: • description of the presented systems architecture and of the processing algorithm implementation, • decisions concerning the design and hardware-software partitioning. www.intechopen.com 2 Ultrasound Imaging 2. Digital signal processing of the ultrasound echoes Diagnostic medical ultrasound devices utilize ultrasound waves at the frequency of 1-15 MHz. The transmitted ultrasound wave, mainly in the form of a train of short pulses, propagates in the tissue where the structures are reflected and returned to the head as echoes. The signal of echoes is initially amplified and filtered in an analogue chain and next, digitally processed using ADC (Analog-Digital Converter) with 8-14 bit of resolution (Thomenius, 2006). The received high frequency signal (called RF – radio frequency) of the echoes is amplitude- and phase modulated carrier frequency signal. The signal is demodulated in the device to obtain baseband frequency. The demodulated echo signal is further processed, depending on the application (Ali et al., 2008). TX Pulser I/Q signal RF signal TX/RX switch VGA A/D Convetrer Demodulation Imaging/Doppler processing Post-processing & Display US Head Fig. 1. Simplified block diagram of an ultrasound diagnostic device The chain of ultrasound signal processing (Fig. 1) is much the same as the chain used in digital telecommunication. Therefore, in both cases, similar system solutions and processing methods are applied. For adaptation to quickly changing processing algorithms used in telecommunication, the Software-Defined Radio (SDR) was developed, in which RF signals are processed using software (Reed, 2002). The generic architecture of SDR systems composed of GPP (General Purpose Processor), DSP (Digital Signal Processor), FPGA (Field Programmable Gate Array) and high frequency front-end blocks was proposed by (Bassam et al., 2009). According to the authors, universal systems of RF signal processing can be implemented based on this architecture. The most popular example of an SDR platform is the GNU Radio project, and the compatible Universal Software Radio Peripheral (USRP) module, designed by a group of enthusiasts, later used in multiple designs and studies. The USRP module consists of a set of broadband ADC and DAC connected to the FPGA, which in turn communicates with the PC via USB interface. The board is prepared to host analogue transmitter and receiver modules, which tailor the solution to specific applications. The FPGA is responsible for signal modulation and demodulation and stream data to and from PC. The developed digital signal processing in C language with interface to Python script language enables quick and easy prototyping of radio signal processing algorithms. With its universal design and availability of a relatively cheap hardware module, the system is quite popular. The project website (http://gnuradio.org) contains information on the application of the project, as well as protocols and telecommunication standards including WiFi, RFID, DVB, DAB, GSM and DECT. The newer and more advanced version of SDR platform named SORA was developed by Microsoft® Research. It has a significantly larger FPGA (Xilinx Virtex 5) chip and a faster www.intechopen.com Hardware-Software Partitioning of Digital Signal Processing in Ultrasound Medical Devices a Case Study 3 PCI-e x8 interface. Moreover, a software stack for multicore processors containing drivers and processing libraries, optimized for efficiency and latency minimization was developed (Tan et al., 2009). The authors also presented a demonstrative implementation of SoftWiFi supporting the 802.11a/b/g communication protocol. The reader will probably notice a very high similarity between the USRP solution and coder-digitizer module of the uScan system described later in the paper. It definitely indicates a similarity between the system solutions and the architecture of SDR and ultrasound devices. Interestingly, the author has not found any report on using SDR solutions for ultrasound RF signal processing. 3. Digital signal processing systems Nowadays, there is a vast number and diversity of digital signal processing systems and methods. Therefore, it is extremely difficult to select an optimal system solution for signal processing of suitable processing power, data throughput and power consumption at a reasonable price. If we limit our choice to the solutions commonly applied in ultrasonography, these will include: • hardware processing – programmable logic systems (FPGA) and ASIC (Application Specific Integrated Circuits), • software processing - general purpose processors (GPP), digital signal processors (DSP) and graphic processors (GPU). Recently there is a tendency to increase software processing because of the development of processor systems and significantly easier implementation process and algorithm debugging as compared to hardware implementations. This tendency has resulted in the development of ultrasound medical devices, requiring a more efficient digital processing and working in a real time regime. The main advantage of software solutions is easy code modification and simpler testing and debugging methods. Implementations in high level languages (e.g. C language) can be easily simulated and verified on different platforms before being implemented on the target platform. Nowadays, the majority of DSP are programmed in C language and have more advanced tools for development. Thanks to their computing power, easy high-level code development, DSP are the core of processing chain in modern devices. There is an alternative trend in software processing, namely general purpose processors (GPP). Modern PC processors have high computing power and special parallel executive units for multimedia processing. These units are in fact separate vector processors, optimized for digital signal processing. Adequate use of parallel processing and multimedia units can replace several DSP working at the same time. As PCs are frequently part of these devices and are used both as controller and user interface, the implementation of processing algorithms using PC seems natural. Real time application of operational systems, such as Microsoft® Windows is problematic as these are not real time systems and therefore they do not guarantee adequate time precision and the required operations being performed in the required time period. Recently, graphic processor units (GPU) are used for signal processing; they are traditionally used for 3D graphics rendering in PCs. The quick development of computer graphics resulted in the development of graphic processors with new capabilities. From specialized 3D processors, GPU have been developed into versatile programmable vector processors. Thanks to its massive parallel internal architecture and advanced memory www.intechopen.com 4 Ultrasound Imaging interfaces, the floating point processing power of GPUs can be even 100 times higher than the power of the fastest PC processors. The new tools for graphic processor programming allow to utilize new computing power, not only for graphic purposes, but also in other, more general applications (Owens et al., 2005). Hardware implementation of digital signal processing is presently dominated by programmable logic systems (FPGA). Thanks to the availability of cheap FPGA versions, they are used not only in small scale production, but also in products released onto the consumer market. The possibility of reprogramming within the system (also during operation), a very wide range of system density (from thousands to millions of gates) and increasingly advanced design tools (McDonald, 2009) contribute to the growing popularity of FPGA systems. Hardware methods offer the highest potential processing speed (Sirowy & Forin, 2008), however the implementation costs are also high, probably the highest. Therefore, hardware and software methods are frequently applied as complementary methods to achieve the required efficiency and versatility at minimal costs. The discussion presented above indicates that digital processing algorithms can be implemented in different ways, based on different solutions. The optimal use of various computational tools and system architecture providing development opportunities is a complex problem from the point of view of science and engineering. 4. Ultrasound devices - state of the art The combined architecture of hardware-software processing devices is currently the main feature of various digital signal processing devices, requiring high processing speed and characterized by a high bandwidth of input data stream. The combined architecture incorporates hardware systems (FPGA/ASIC), responsible for initial processing and aggregation and/or distribution and further software processing systems (DSP, GPP, GPU processors). We can see a clear tendency to extend the scope of software processing and minimize hardware solutions. The popularity of software solutions results not only from easier implementation, testing and debugging, but also from the possibility of adaptation of these devices to quickly changing market requirements and new algorithms. On the other hand, the state of the art FPGA systems should be now more frequently treated as programmable SOC (System on Chip) systems rather than simple gate arrays. The availability of new tools for algorithm implementation in FPGA (e.g. Xilinx® AccelDSP, Altera® DSP Builder) significantly facilitates the design and transfer of complex computational algorithms without the need to encode them in hardware description languages. The available IP (Intellectual Property) processor blocks (e.g. Xilinx® Microblaze, Altera® NIOS, Actel® Cortex-M) enable the creation of single- and multiprocessor systems directly in FPGA structures. The processor blocks support the implementation of extended instruction sets using hardware solutions built from the surrounding logic (the so called instruction accelerators). Alter® C2H tool is an example of such solution that compiles selected code fragments in C language directly for hardware implementation, warranting 15-73-fold speedup for selected functions (Altera, 2006). Integrated systems based on the reprogrammable FPGA were described by (Garcia et al., 2006). The model example of combined ultrasound system architecture is a Doppler device developed by (Ricci et al., 2008). The construction of this system is much the same as that of digiTDS system presented below. The module designed by Ricci et al. is composed of FPGA (Altera® Stratix) executing hardware processing functions, and DSP (Texas Instruments® www.intechopen.com Hardware-Software Partitioning of Digital Signal Processing in Ultrasound Medical Devices a Case Study 5 TMS320C67). The module communicates with a PC and is responsible for the control and display functions through the USB interface. The FPGA has implemented a quadrature demodulator, a set of digital filters and a decimator. The initially demodulated and decimated data stream is transferred to the DSP, executing standard Doppler algorithms. Although the reported system is not a commercial solution, its construction is similar to that of the devices offered on the market. Another device, an ultrasonograph platform ULA-OP, was designed by the same group (Tortoli et al., 2009). Although it is a 64-channel system, it is based on the same general concept. The basic difference is that this system utilizes multiple FPGA chips connected to multichannel ADC and a central hub FPGA, responsible for data aggregation (in total: 5 FPGA Altera® Stratix systems, each having its own 512 MB of DDR memory). The efficient DSP (Texas Instruments® TMS320C6455) is responsible for software processing. The system is connected to PC via USB port and the PC controls the system and displays the results. As we can see, although it has many channels, the general scheme of device design does not differ from that of a dual channel Doppler system. The next system described in literature is a hand-held 32-channel ultrasonograph, developed by (Lee et al., 2009). Its architecture resembles the ULA-OP system. The main differences are that the ultrasonograph uses a single large FPGA (Xilinx® Virtex 4) and an additional application processor (Intel® XScale PXA270), executing the user interface and display on the integrated LCD panel. The system is completely independent (does not require connection to PC) and optimized for the size and power consumption. The distribution of processing tasks between the hardware and software is in conformity with previous solutions. 1024-channel SARUS (Jensen et al., 2007) system is a unique ultrasound research platform. Unlike the previously described solutions, the system is mainly based on the on-line hardware and off-line software processing. It consists of 64 identical DAUP (Digital Acquisition and Ultrasound Processing) boards, each operating 16 transmitting-receiving channels and an external computer cluster with a disk array. Each of the DAUP boards contains 4 FPGA (Xilinx® Virtex XC4VFX100) with a dedicated 1-4 GB DDR2 memory and one control FPGA (Xilinx® Virtex XC4VFX12). Technical solutions and the potential of SARUS system are overwhelming, however, this is only a research system and its architecture does not allow it to be directly adapted to commercial systems. The hardware only processing chain for ultrasound system was presented by (Chang et al., 2009). A single-channel acquisition and processing system consists of ADC (Analogue Devices® AD9430) connected to FPGA (Altera® Stratix EP1S60F1020C6), executing the entire processing. Due to the planned application of the above system in studies on mice whose heart rate reaches 13 beats per second, the authors made an attempt to perform imaging with the speed of 400 images per second. The application of simplified and optimized processing algorithms (including envelope detector) allowed to obtain the designed processing speed. The reconstructed images were next transmitted to PC via PCI interface and displayed using a graphics card with the implemented scan-converter. The presented solution indicates how high processing speed can be obtained in systems using hardware processing. A completely different solution is a C-scan imaging system presented by (Fuller et al., 2009). It has a custom built ultrasonic array transducer, composed of 3600 elements with an integrated transmission-reception electronics. The echoes received from selected depth are then transmitted to a DSP (Analog Devices® Blackfin BF561) which executes the entire www.intechopen.com 6 Ultrasound Imaging digital signal processing and display on the built-in LCD monitor screen. The obtained display speed of 43 images per second for the C-scan indicates that in some applications, pure software processing is sufficient to obtain a practical outcome. Interestingly, ready-to–use solutions for commercial ultrasound scanners are offered by Texas Instruments® company, the leading manufacturer of DSP. Texas Instruments® has developed an optimized function library for applications in ultrasonography using the offered TMS320C64 series DSPs (Thomas, 2010). Moreover, (Pailoor & Pradhan, 2008) present specific architectures and solutions for portable medical ultrasonography systems. These are systems with combined hardware-software processing. As in the previous presented systems, the first processing stage (beamforming and demodulation) is executed by hardware, while the subsequent stages are executed by DSP. The TI offers a complete development environment and integrated circuits portfolio to build almost the complete systems (TI doesn’t offer FPGA). The final solution presented in this paper is a cheap ultrasonograph with a single processing channel in the form of an integrated head connected to PC via USB port (Richard et al., 2008). The entire simplified electronics is integrated on a miniature board, placed directly in the ultrasound head. The RF echo signal undergoes analogue demodulation and it is digitally processed and transmitted to PC via USB interface. An analogue demodulation of RF signals is no longer used in the devices; in the presented case its application results from construction simplification. Due to low power consumption, the head is supplied directly from PC via USB. The whole processing and visualization were implemented as software for PC. The quality of an ultrasound image fails to meet contemporary standards, however, on the other hand, its price is unmatched. The concept of system solution, minimization of electronics and total PC processing is similar to that of the uScan described further in the paper. 5. Case study Both systems are based on a mixed model of hardware-software signal processing. The systems of RF echo signal acquisition are based on a fast A/D converter and FPGA, being an interface to the further part of the system. The application of FPGA at the beginning of the digital chains is now a standard solution. Apart from signal acquisition, the FPGA most often executes hardware processing of a digital signal, whose goal is to reduce the data stream for further processing. Next, the processing is executed by DSP, GPP, GPU or combinations of these processors. 5.1 High frequency ultrasound scanner - uScan High frequency ultrasonography is an imaging method using ultrasound waves of more than 20 MHz frequency. Due to the obtained high resolution and low penetration depth of imaging, this method is applied in dermatology, ophthalmology, cosmetology and for examining small animals. uScan system is a high frequency ultrasound scanner (20-40 MHz) with the implemented coded transmission/excitation function. The coded transmission method involves transmission of long series of impulses and subsequent time compression of the received echoes to restore resolution. With the long bursts, it is possible to obtain a better signal-tonoise ratio, and the resulting image contrast improvement. Unfortunately, the echo time www.intechopen.com Hardware-Software Partitioning of Digital Signal Processing in Ultrasound Medical Devices a Case Study 7 compression algorithm, which is in fact, matched filter to the transmitted signal, requires a very high computing power. The device is based on the concept of absolute minimization of electronics and computation based solely on CPU and GPU. The electronic module of coder-digitizer (Fig. 2) executing all the transmission and reception functions is 90 mm × 90 mm in size and is based on a single low-cost FPGA chip (Xilinx® Spartan 3 XC3S200). The transmission involves the use of an arbitrary waveform generator with a 14-bit DAC (Analogue Devices® AD9744) operating at 200 MSPS speed. The acquisition is performed using a 12-bit DAC (Maxim® MAX1214) also working at a speed of 200 MSPS. The digital RF echo signal is transmitted to PC via USB 2.0 interface (Cypress Semiconductor® CY7C68013A). The processing software executes the function of time compression on a CPU processor. The sector geometry conversion function (scan-converter) required to obtain a correct display of an image generated by the sector head is executed by a graphics processor. FPGA Xilinx DAC 1 Power Amplifier ADC Input Amplifier DAC 2 TGC Head USB 2.0 interface USB PC Fig. 2. Block diagram of the coder-digitizer module (left), and the board photo of size 90 mm x 90 mm (right) The module was designed for stream data transmission; therefore it contains only a buffer in FPGA for a single RF signal line. The line data have to leave the buffer before the next acquisition, i.e. they have to be sent to the USB controller system. The USB system contains 4 buffers 512 bytes each, which are used for communication in the ping-pong mode. A typical length of a RF line is 2048 samples, which gives 4096 bytes of data (two bytes per sample). On an ACER TravelMate 4650 notebook (Intel® Pentium M processor with a clock 1.7 GHz, 512 MB of RAM), the USB transmission throughput obtained in the test mode was close to 40 MB/s. However, in real conditions, when signal line acquisition was released periodically by ultrasound head movements, buffer overflow errors were observed. This indicates minor transmission interruption, i.e. changes in instantaneous data reception speed by PC. It was concluded that the “interruptions” occurred at system driver level. Microsoft® Windows does not guarantee the time limit of event and interruptions handling from the input/output devices. The overflow errors result in image loss and entail the necessity of subsequent image acquisition synchronization with the extreme position of the head movement. As the coder-digitizer module featured no larger data buffer, it was necessary to limit the head movement speed (and thus image display) to 10 images per www.intechopen.com 8 Ultrasound Imaging second, which is still a sufficient speed for still structure imaging. An alternative solution would be to number the transmitted lines and an additional image filling algorithm with the obtained complete lines. Incomplete or deleted lines of the image would be filled by the fragments of previous lines or by interline interpolation. The digital RF signal processing functions were implemented as a whole in the software domain and distributed between the CPU of the PC and graphic card processor depending on their specificity (Fig. 3). The code compression and detection algorithm is executed on CPU, while the scan-converter and additional post-processing functions are executed on GPU. Such architecture enables execution of more demanding processing systems without the need to employ specialized and expensive multiprocessor systems. Coder-digitizer module Image display Code compression Scan conversion Envelope detection Image filtering CPU Processing GPU Processing Digital RF Software RF processing Fig. 3. Data flow of the software RF digital signal processing Fig. 4. Photo of the complete uScan system Software efficiency tests for processing and visualization were performed on ACER TravelMate 4650 notebook in configuration: Intel® Pentium M processor with 1.7 GHz clock, 2 MB second level cache, 512 MB RAM, graphic processor Nvidia GeForce Go 6600 with www.intechopen.com Hardware-Software Partitioning of Digital Signal Processing in Ultrasound Medical Devices a Case Study 9 64 MB of video memory. For an ultrasound head scanning with the speed of 7 images/s, the CPU load was 37%. When extrapolating the processing time, the system was found to be able to run at the speed up to 25 images/s in this configuration. The obtained solution meets the presented requirements for processing speed and image display in a real time regime (up to 10 frames per second). The applied balanced architecture of digital processing with task distribution between CPU and GPU ensures very high processing efficiency with simultaneous limitation of electronic systems and hardware processing. The limitation of this system is the speed of data transfer via USB interface, as the computing power would allow for significantly faster display. The quick development and increase of CPU/GPU system efficiency ensures the possibility of increasingly complex processing algorithms without the need of hardware replacement. 5.2 Transcranial Doppler System - digiTDS The pulse wave Doppler method is applied as a standard for intracranial diagnosing allowing to measure the flow at selected depth. The state of the art diagnostic solutions are mainly focused on complex assessment and monitoring of cerebrovascular flow in different pathological conditions, especially those potentially affecting normal function of the central nervous system. The digiTDS device is a multigate transcranial Doppler system. This system is composed of two electronic modules of 130 mm × 82 mm dimensions (Fig. 5), responsible for high frequency signal transmission, acquisition and demodulation, and the PC responsible for Doppler signal processing after demodulation and data presentation. Fig. 6 presents the Fig. 5. Photos of the digiTDS’s electronic modules: digital module (left), mixed-signal module (right) www.intechopen.com 10 Ultrasound Imaging RF input TGC ctrl TX circuit ADC 14-bit Power Supply FPGA Altera Cyclone 3 TGC DAC TX DAC Config PPI 128MB SDRAM 4MB FLASH SPI EBIU Generator low jitter Prog EBIU DSP Blackfin 537 +5V +12V USB controler USB 2.0 Ethernet 10/100TX UART RS-232 SPORT Stereo CODEC PC JTAG Boot DataFlash SPI Audio out SD Card Fig. 6. Simplified block diagram of the digiTDS system system block scheme. On an analogue-digital module board, there are two independent transmission-reception channels (the device supports two heads operation, the so called bilateral operation). A 10-bit DAC (Texas Instruments® DAC5652A) operating at a speed of 64 MSPS is used to generate the transmitted signal. The reception track contains amplifiers with the function of time gain compensation (Analogue Devices® AD8331) and a dual 14-bit ADC (Analogue Devices® AD9640) with sampling frequency of 64 MHz. The digital signals from/to converters are connected to FPGA on the digital module board via intermodule connector. The digital module consists of FPGA (Altera® Cyclone 3 EP3C25), DSP (Analog Devices® Blackfin BF537) as well as USB device controller (PLX® NET2272) and an Ethernet interface. Unlike the uScan, digiTDS may run independently of PC, thanks to its own control resources - DSP processor. The system communicates with PC via USB 2.0 interface. The digital processing chain consists of hardware processing in the FPGA and software processing using DSP and PC’s CPU. The application of hardware resources during early processing stage allowed to significantly decrease the required data throughput and computing load of PC. It also allowed to use a low-power single-board PC with Intel® ATOM processor and to extend the operating time on batteries. The distribution and implementation of processing tasks were performed on a constructed device prototype. Especially moving the realization of filtration in multiple gates from the software to FPGA enabled significant processor time saving. Fig. 7 presents the Doppler digital signal processing track on digiTDS device with task distribution between the FPGA system and software. www.intechopen.com