Lowering the cost of medical-imaging R&D

Medical-equipment companies are shifting from a vertical-integration model to a system-integration model to do more with less research and development.

Allan Evans, Samplify Systems Inc -- EDN, February 3, 2011

A look inside a typical piece of medical-imaging equipment reveals myriad technologies from many engineering disciplines that must collaborate to form a system. For example, developing X-ray scintillators in CT (computerized- tomography) machines to convert X-rayparticle energy into photon energy requires materials-science research into rare-earth materials. Capturing this photon energy over wide dynamic ranges requires custom analog-frontend electronics, and acquiring the data from tens or even hundreds of thousands of such X-ray detectors requires customized storage subsystems and interface technologies. Complex software algorithms then form images from this substantial volume of raw data. Radiologists, sonographers, and other medical-imaging professionals examine these images to make their diagnoses. This vertically integrated, multidisciplinary engineering from materials-science research to custom silicon and complex software design has historically required hefty R&D budgets, which only large companies could afford.

The current economic environment has caused an increased focus on price and performance and forced medical-equipment OEMs to re-evaluate their R&D budgets to focus on technologies that are truly keys to their competitive advantage, outsourcing other nondifferentiated imaging functions to outside semiconductor and subsystem suppliers. These merchant suppliers of technologies can, in turn, enable cost reductions by amortizing their R&D costs over the entire industry, not just across the market share of one equipment manufacturer.

The evolution of the ultrasound-equipment market provides an instructive example of how this trend will play out. Figure 1 depicts a 10-year-old ultrasound machine whose probe contains 64 to 256 piezoelectric transducers, each individually connected through a microcoaxial cable to the console. The probe required a specialized manufacturing process for matching gains and delays across elements. As manufacturers introduced imaging techniques, analog front ends primarily comprised discrete components. Top-tier imagingequipment manufacturers held this “black art” close to the vest. Processing the tens or hundreds of gigabits per second of data could take place only in custom ASICs. When medical applications began to use analog CRTs for ultrasound displays, manufacturers began to develop proprietary interfaces and backplanes for image processing.

Now fast-forward 10 years (Figure 2). The high-level picture looks the same, but the components are different. Manufacturers now typically outsource the labor-intensive manufacturing of the probe to ODMs (original-design manufacturers). As the imaging modes matured, semiconductor companies began to understand the specifications for analog performance. Consequently, they began to introduce highly integrated analog front ends and modules with as many as 32 channels for the ultrasound-equipment market. Since mid- 2008, companies have introduced more than a dozen analogand mixed-signal semiconductor devices for this market.

Digital signal processing for functions such as beam forming is also moving to the merchant semiconductor suppliers. Previously, at the 180-nm node, toptier ultrasound OEMs could justify the NRE (nonrecurring-engineering) costs for custom ASIC designs for digital processing. Many of these designs also featured mixed-signal integration with ADCs at 8 or 10 bits/sample. However, with new imaging technologies, such as tissue harmonic imaging and pulsemode color Doppler, leading-edge ASIC designs have become cost-prohibitive, and the integration of high-performance 12-bit ADCs has become too complex. Today’s mask costs, even at the 65-nm node, are too expensive even for market leaders. As a result, ultrasound OEMs have turned to FPGAs for their DSP functions. Manufacturers are building FPGAs at the 45- or 40-nm node and are announcing 28-nm parts, and they feature gate densities many times larger than ASIC densities at the 180-nm node. Now ultrasound OEMs can purchase an entire analog-front-end subsystem complete with beam forming, which can generate ultrasound images out of the box, rather than taking six months and many hundreds of thousands of dollars to custom-design their own.

The range of applications for ultrasound equipment has increased greatly by enabling OEMs to focus their human and capital resources on their core features. Specialized ultrasound machines are now available for mammography, surgical guidance, and cardiography, and new imaging form factors are emerging for emergency-room and ambulance applications. The availability of these products has lowered R&D costs, allowing OEMs to enter the ultrasound-equipment market with a focus on niches. Top-tier OEMs also have benefited from these subsystems and silicon by allowing these OEMs to rapidly broaden their product portfolios rather than rely on a one-size-fits-all machine. The health-care industry has benefited through an improved price-to-performance ratio across a range of applications.

These approaches can provide the same benefits to imaging technologies besides ultrasound. For example, a CT machine is the result of vertically integrated development, much the same as ultrasound was 10 years ago. Again, OEMs’ research into materials science results in customized X-ray scintillators that turn X-ray-particle energy into photons. These scintillators require customized analog-front-end silicon. Customized algorithms and protocols amplify, digitize, encode, and aggregate these scintillators’ data and then transmit it to a workstation across a slip-ring device. In the CT-image-reconstruction workstation, customized network-interface cards terminate custom protocols, and custom RAID (redundantarray-of-independent-disk) controllers buffer the raw data at high rates until the workstation can reconstruct the image at a lower rate (Figure 3). These workstations can often reconstruct a 30-second CT scan in two to five minutes.

The high data rates for real-time CT data acquisition drive much of the need for custom hardware in the workstation in which the data is stored. Consider a 64-slice CT machine with 1000 detectors per slice with a rate of 10k samples/sec and 16 bits/sample. The aggregate throughput requirement is 1280 Mbytes/sec, a rate that exceeds the throughput of commercially available RAID controllers, which top out at 800 Mbytes/sec. CT-signal compression can reduce the throughput requirement of RAID controllers. For example, GE Healthcare, Samplify Systems, and Stanford University have demonstrated the efficacy of signal compression in both lossless and nearly lossless modes to achieve compression ratios that provide compelling reductions of 3-to-1 or even 4-to-1 in data-acquisition rates (Reference 1).

Figure 4 shows an original image and a sample image of 3-to-1 compressed data. For more than 400 images, a Stanford radiologist could not distinguish the image of compressed samples from the image of noncompressed samples. With signal compression on the rotor side of the slip ring and decompression in software on the workstation, the CT system realizes the benefit of bit-rate reduction across the entire signal chain, including the slip ring, the network-interface card, and the RAID controller. Integration of signal compression into CT slip rings makes the compression transparent to the rest of the CT system so that such compressing slip rings can be drop-in replacements for older slip rings.

In CT machines, image reconstruction typically occurs three to 10 times slower than does CT data acquisition. This asymmetry creates the need for RAIDs in the system to buffer the raw data until the CT machine can reconstruct it into an image. Because the RAID is the last link in the data- acquisition chain, CT-machine designers must build the workstation on the CT gantry by considering data-acquisition rates instead of image-reconstruction rates. A new architecture helps designers build CT workstations based on the lower image-reconstruction rates rather than the higher data-acquisition rates (Figure 5).

With compression of the X-ray-detector data in the slip ring, you can now replace a custom multilane fiber-optic interface with a standard storage interface, such as FibreChannel or InfiniBand. This approach allows you to use off-theshelf RAIDs outside the workstation. In this architecture, an off-the-shelf FibreChannel or InfiniBand switch provides the connectivity between the gantry or the data-acquisition subsystem, storage subsystem, and workstation. The workstation can now use standard low-cost network-interface cards rather than the customized interfaces that today’s gantry-to-console connectors use.

Furthermore, because image-reconstruction rates are several times lower than data-acquisition rates, designers can build decompression into the RAID, enabling the integration of compression into CT systems in a manner that is transparent to image-reconstruction software. With the elimination of the need for a RAID or its transition into the gantry, the designer can now base the workstation on commodity PC-server hardware.

With R&D budgets tightening, medical-imaging OEMs must deliver next-generation technologies across a wider range of market segments with lower development costs. This move requires them to rationalize technologies that target the competitive advantages of their machines.

---

References

Wegener, Albert; Naveen Chandra; Yi Ling; Robert Senzig; and Robert Herfkens, “Real-time compression of raw computed tomography data: technology, architecture, and benefits,” Proceedings of the SPIE Medical Imaging Conference, Volume 7258, pg 7258H, March 2009.

---