Medical imaging applications, such as CT scanning, move extremely large amounts of data through a system at a very high rate. In order to avoid bottlenecks at certain points in the process and to alleviate storage space concerns, medical device manufacturers of this type of equipment are looking for compression solutions. This article reviews one potential response to the problem.

Modern CT scanners digitize tens of thousands of x-ray channels at sample rates in the range of 1 to 7 ksamp/sec with 16 to 24 bits per sample, resulting in an aggregate bandwidth of many Gbps across the slip ring and to disk drives for subsequent storage. As the industry strives for increased spatial and temporal resolution and as newer CT scanners utilize multiple x-ray sources that require a multiplicative increase in sensor count, the bandwidth demands on CT slip rings and disk storage subsystems will soon exceed 10 Gbps.

Figure 1: CT bottlenecks (modified from US patent 6,862,299)

Figure 1 illustrates the two primary bottlenecks for raw CT samples. The first bottleneck occurs when the raw x-ray sensor signals (digitized by tens of thousands of A/D converters) are sent across the slip ring. The second bottleneck occurs as the data acquisition computer transfers the samples to a disk array for storage. In most CT applications, image reconstruction is performed after the data is acquired because reconstruction speeds usually cannot keep up with data acquisition rates. Because the CT community continues to desire faster scan rates, higher resolution, and wider coverage per rotation, more A/D converters will be used in the future, thus exacerbating the bandwidth and storage bottlenecks both at the slip ring and at the disk array.
Resolution, Slices, Views, and Rotation Rates
Most third-generation CT scanners use arrays of A/D converters called “slices.” CT manufacturers such as GE Healthcare, Siemens, Philips, and Toshiba offer CT scanners with between six and 64 slices, where each “slice” typically contains between 512 and 1024 x-ray sensors. Each x-ray sensor is attached to an A/D converter that measures several thousand raw x-ray counts per second. Because x-ray counts have a very wide dynamic range, A/D converters that digitize x-ray sensors can require up to 24 bits of resolution, depending on vendor implementation. In some instances, the full range of the A/D converter is binned into a smaller number of “buckets,” thus reducing the average number of bits per sample to 16 or less.

Figure 2: Typical 1024-sensor raw CT signal
Figure 2 shows a raw CT signal from a 1024-sensor array. In a high-end machine, up to 64 of such signals are combined into a “view,” where the x-ray sensor values represent one angular offset of the rotating CT gantry. Gantry rotation rates of two to three rotations/second are common, with angular resolutions of 0.3 degrees per view. The following equation combines all of these values to generate the prodigious flood of sampled data generated by modern CT scanners:

16 bits/sample x 1024 sensors/slice x 64 slices/view x 1000 views/rotation x 3 rotations/sec = 3.14 Gbps

Compression for Medical Imaging
One may ask why compression of raw sensor data hasn’t been used before for medical imaging applications (e.g., CT, ultrasound, MRI, PET, etc.) The short answer is simple: no high-speed lossless compression algorithm has been available that achieves acceptable (~2:1) compression ratios for the raw sensor data from which medical images are formed. Because the popular compression algorithms for consumer audio, speech, images, and video are all lossy algorithms, they cannot deliver the lossless results required for medical imaging. Also, audio and video compression algorithms achieve their impressive 10:1 and 20:1 compression ratios by knowing how human hearing and vision operate. With this knowledge, compression algorithms can remove the bits that won’t be noticed during audio or video playback. These auditory and visual weaknesses cannot be exploited for raw signals, since these signals are not converted to audible or (directly) visible images. In CT instead, the raw signals are processed using a sophisticated and often proprietary set of algorithms (inverse Radon transform, beamforming) that generate the patient images.
Real-Time Compression
In response to these relentless demands for increased CT bandwidth and storage, there now exists a real-time compression technology with the goal of alleviating the CT scanner’s slip ring and disk drive bottlenecks. The first algorithm to support real-time compression that is effective for raw CT sensor signals, and at modest complexity, now exists. By compressing the raw, digitized x-ray A/D converter samples on the gantry and transferring compressed packets instead of raw samples to the disk drive array in the image reconstruction PC, compression reduces both CT scanner bottlenecks at once.

Like all medical devices, the Food and Drug Administration (FDA) approves CT scanners. To use compression in the current generation of CT scanners, only lossless compression of raw CT scanner data can be employed. With lossless compression, the signal that results after decompression is identical (bit for bit) to the original signal. Just as WinZIP or PKzip achieve lossless compression on computer files, new compression techniques operate on sampled data systems. However, these compression techniques achieve much better compression ratios on CT sampled data than WinZIP or PKzip can achieve. In addition to higher compression rates, lossless compression has much lower computational complexity than competing algorithms, enabling compression to operate approximately 10 times faster than either WinZIP or PKzip. Lossless compression ratios of between 1.8:1 and 2.2:1 are achieved on a variety of CT raw sensor data captured from various sections of the body (head and neck, shoulder, abdomen, hip, and various resolution phantoms).

As evidenced by the widespread adoption of lossy compression algorithms for audio, speech, video, and image signals, lossy compression for CT scanner signals that offers higher compression ratios may eventually be accepted (and FDA-approved) for CT scanners. As the visual perceptibility of lossy compression of raw CT signals becomes better understood, the higher compression ratios offered by lossy techniques may join existing lossless compression methods as useful new CT tools. The next step is to develop appropriate image quality metrics that quantify and minimize the effects of lossy compression on raw CT signals. The goal of these efforts is to obtain the highest compression possible while maintaining a minimum fixed quality threshold that is selectable by the equipment manufacturer or perhaps even by the end user.
System Integration and Compression Complexity

Figure 3: CT scanner with integrated compression

As illustrated in Figure 3, compression and decompression can be smoothly integrated into existing CT architectures. Both lossless and fixed quality compression algorithms are already available in popular field programmable gate arrays (FPGAs) from such top providers as Altera and Xilinx. FPGAs are a significant part of the signal processing chain in most CT scanner gantries, and low complexity (1,800 Xilinx slices or 3,000 Altera LEs) allows it to be integrated into existing CT FPGAs. Decompression of the raw CT signals can be implemented in a variety of ways, using a dedicated ASIC or FPGA, or in one of the processing elements currently used for image reconstruction in the imaging PC. The particular choice of where decompression is performed is system-dependent. The advanced decompression technology is available in whatever form is required: dedicated hardware, intellectual property (IP), or software.
What’s Next?
In the future, integrated data acquisition ICs that incorporate compression may become an integral part of tomorrow’s CT scanners and other medical imaging devices. By integrating compression as an intellectual property block into A/D converters for medical imaging, the package size, pin count, and power consumption of such A/D converters can be decreased. As mentioned earlier, CT scanners may be able to further increase the bandwidth across the slip ring by applying fixed quality compression in the future. Before using lossy compression, CT researchers will have to demonstrate that the image quality is not degraded for its diagnostic purpose.
CT scanner manufacturers will benefit by integrating compression of raw x-ray sensor data. The benefits include increased slip ring bandwidth that supports a 2x increase in the number of x-ray sensors, or the rate at which the sensors are read. In addition, the bandwidth and storage required to store the raw CT sensor data is also decreased by 2x. The new algorithm achieves approximately 2x lossless compression on raw CT signals and can be smoothly integrated into the FPGAs and microprocessors that are already present in CT scanners.OnlineFor additional information on the technologies and products discussed in this article, visit the following websites:
  • Al Wegener is the chief technology officer and founder of Samplify Systems, a venture-funded Silicon Valley start-up whose patented compression solutions reduce bandwidth and storage bottlenecks in sampled data systems. Wegener is a DSP engineer, technical manager, and inventor with more than 25 years of experience in defense electronics, professional and consumer audio, and wireless applications. He can be reached at