RTC Magazine

Until recently, many inspection and control systems have required a human in the loop to provide the actual inspection. However, due to the inefficiency, and sometimes ineffectiveness, of manual inspection, automated inspection and control has become the most sought-after method. This has proven to be no easy task, especially in the complex world of flow cytometry and cell sorting. This is the process used to measure the physical or chemical characteristics of a biological cell and, using this information, sort cells from a sample. Such high-end applications require an intelligent system that can also provide a high level of flexibility and reprogrammability, so that the system is, to some extent, a general-purpose one that can be tailored to the needs of individual applications.

For years, DSPs have been the processor of choice for applications requiring the type of compute-intensive, high-speed calculations used in both medical and bioscience fields, as well as in manufacturing. FPGAs are now being touted as do-everything processors, especially when reconfigurability is a requirement. Often, the best solution is a combination of FPGAs and DSPs working in conjunction with each other (Figure 1).

Basic Inspection and Control System Needs

All systems used for inspection and control, regardless of industry, include two stages of processing. The first stage is the initial inspection during which the system acknowledges the presence of a sample. In the second stage, the required feature extraction is performed. This consists of gathering certain characteristics from the unit under test, often using digitized images for pattern recognition and feature extraction; providing the “intelligent” processing of these images; and ultimately determining if this is a “good” sample or a “bad” sample.

The processing required for the initial inspection, although fairly minimal, needs to be repeated indefinitely as the system consistently monitors for the next unit. Regardless of the type of control and inspection system, the second stage of processing is quite complex.

Flow Cytometry and Cell Sorting Systems

In high-end bioscience applications, automated inspection and control is often required to perform tasks that would otherwise be impossible, since they cannot be done manually. The sorting and classification of biological cells via flow cytometry and cell sorting is one such area. This type of control and inspection system is very similar to the basic system described above in that there are two stages: the initial constant inspection to determine the presence of a cell, and the intelligent processing. The end goal of the system is to divert a droplet containing a cell into the appropriate receptacle by analyzing the cell’s identifying characteristics.

Before each droplet is released from the sample, the system must identify that a cell is present and begin to analyze it. In order to analyze and isolate certain characteristics of a cell, a fluorescent compound is formulated specifically to highlight those characteristics. When added to the sample, this compound binds itself to the cells that possess the desired characteristic. The pressurized cell suspension is forced into a tube and then through a needle, making the cells form a single-file line. The line of cells is then forced through a small nozzle, emerging as a jetting stream. The jetting cells pass one at a time through a laser beam where the cells scatter laser light.

If the desired characteristic is present, the compound in the cell also emits fluorescent light. This fluorescence is detected by a photo sensor, which converts the fluorescent light into a momentary voltage pulse. This stream of voltage pulses is fed into an A/D converter that continuously samples the stream of pulses. Within microseconds, the system needs to perform the pulse detect, identifying that a cell is present, then extract the desired features from the pulse, analyze these features and ultimately determine whether or not the cell is to be sorted.

These tasks take place in parallel so that the system may, at any single point in time, simultaneously detect one cell, analyze another cell, and route yet another cell. This requires a microsecond-level timeframe. Within one second, there could be up to 100,000 cells shooting out of the system, all needing to be measured, classified and routed. Typically, the sorting decision must be made 100 microseconds after the cell passes through the laser beam.

Flexible, Intelligent Cell Sorting

These flow cytometry and cell sorting systems have traditionally utilized complex, custom analog circuitry that has remained unchanged for many years. One manufacturer, iCyt Visionary Bioscience, decided that it was time for a change. The company wanted to design an intelligent system that would also provide a high level of flexibility and reprogrammability, enabling it to be somewhat general purpose, yet capable of being tailored to the requirements of specific applications.

The older systems sampled only the maximum peak fluorescence coming from a particular cell, providing merely a single data point. The proposed new design would capture at least 100 samples per cell, giving the system more information to analyze, such as rise times, fall times, width and integrated energy. This additional data would greatly reduce false detections and missed detections.

A cell typically passes through the laser beam in about one microsecond. Within no more than 100 microseconds, the droplet containing the cell must be routed to the correct receptacle. During this time, the system must detect the presence of a cell, extract its relevant features, analyze this data and sort the cell. This requires a sampling rate of at least 100 MHz. At 2 bytes/sample, the required sustained throughput is 200 Mbytes/s. The technology chosen had to support this throughput while also providing extended precision and dynamic range on unpredictable and changing data in a fixed amount of time.

Although the proposed system would be targeted at cell sorting, many specialized applications are possible within that field. The system would therefore need to be flexible, enabling not only a variety of uses, but also satisfying requirements that do not yet exist.

A Choice of Processors

When determining which technologies should be utilized in the design, the company’s designers had to take into account four separate requirements: low latency, high throughput, complete flexibility and the system’s ability to be rapidly adapted and its design updated.

The question then became which processing technology to use, from a traditional microcontroller or microprocessor, to a DSP or programmable logic device.

The speed and dynamic range required for the system pointed to a DSP. They deliver sustained, low-latency, high-throughput data processing. The additional dynamic range provided by a floating-point DSP, required by this proposed system design, also ensures that the system will easily support future applications. A general-purpose microprocessor would be hard-pressed to match this set of abilities. Flexibility, unmatched algorithmic capabilities optimized for fast floating-point math, and the benefit of programming using C, effectively decreasing development time, are all benefits of designing with a DSP.

DSP-Only Design vs. Hybrid DSP/FPGA-Design

The prototype used four high-end, floating-point TigerSHARC DSPs per channel to accomplish the first and second stages of processing in the flow cytometry and cell sorting system. These ultra-high-performance, real-time multiprocessing DSPs were chosen for their high throughput, low latency and deterministic characteristics. They provide continuous complex processing at high speeds in a fixed amount of time, crucial to this application.

After further consideration, it was determined that the initial stage of processing could be accomplished much more efficiently with an FPGA. Utilizing a lower-cost FPGA for the pulse detect would save the interesting data for the DSPs, which are particularly well suited to the type of high-throughput, data-stream processing required in the second stage.

The FPGAs, on the other hand, could easily perform the processing required for the initial cell detect. This data does not require complex computational math, but does require fast, repetitive processing. Since the DSPs are not customized for this type of data analysis, they were wasting large amounts of processing effort on the first stage. Although the FPGA processing only accounts for 20% of the system’s total processing, the number of DSPs required was cut in half once the FPGA was added.

Significant design work would be required to ensure that the FPGA and DSPs operated efficiently together. Many design engineers are accustomed to working primarily with either DSPs or FPGAs, but not both, and might be unfamiliar with the requirements of integrating the less familiar silicon into a design. An architecture which provides that for designers becomes crucial. This is especially true if it also enables them to easily change the routing of inputs and outputs of the signal processing system for future expansion and modification.

The final subsystem (Figure 2) came in the form of BittWare’s hybrid embedded design, which uses the company’s custom ATLANTiS architecture to interface between the FPGA and DSPs. This architecture is an I/O routing scheme in which every I/O can be dynamically connected to any other I/O. By employing this architecture and utilizing both DSPs and FPGAs in the design, a powerful, completely flexible system was created. The FPGA provides the initial processing, performing the pulse detect, and passes the data along to the DSPs only when a cell is actually present. The DSPs then provide feature extraction and intense data analysis.

The architecture (Figure 3), implemented in the FPGA, provides one solution for the three dilemmas typically faced in multiprocessor designs. These are: how to allocate the I/O bandwidth among the processors, how to easily connect the various I/Os together while retaining the ability to modify them, and how to integrate the FPGA processing with the DSP processing. It solves all of these dilemmas by enabling communication between the TigerSHARC link ports and all other I/Os connected to the board and by providing continuous 2 Gbyte/s throughput. This communication can be point-to-point, or one input can broadcast to all or multiple outputs. The I/Os can be connected or disconnected from each other as requirements dictate without the need for recompiling or changing cables.

This architecture consists of one or more switches connected to a configuration register that is controlled by the designer using a GUI. The routing can be changed at any point by reprogramming the configuration registers. Since all I/Os connected to the board are input into the FPGA, and thus into the architecture itself, standard and/or custom FPGA processing blocks can be inserted in the queue at any point during the transfer of data. For both novice and seasoned FPGA designers, this architecture enables seamless data transfer and the option of including processing blocks.

The final system using both DSPs and an FPGA, coupled with an architecture that gives designers the ability to easily integrate the two, provides an intelligent and flexible flow cytometry and cell sorting system. The design easily handles the initial system requirements, while also providing enough flexibility for future enhancements and upgrades.

BittWare
Concord, NH.
(603) 226-0404.
[www.bittware.com].