Machine Vision and Medicine – A Healthy Partnership

Machine vision, like the medicine a doctor gives a patient, is helping manufacturers maintain healthy productivity, but that’s not the end of relationship between the medical industry and automated inspection systems. Machine vision is helping to track and inspect medical devices and pharmaceuticals as well as enhancing the performance of medical imaging modalities.

FDA requirements
Machine vision fulfils three general types of applications within the medical industry: product quality inspection, product tracking and medical imaging. The first two applications can vary significantly based on the product under inspection, but they do have one thing in common: FDA requirement CFR211. This federal regulation requires 100 percent inspection and tracking of products destined for use inside the human body.

According to Fabio Perelli, product manager for frame grabbers and stand-alone systems for Matrox Imaging (Dorval, Quebec, Canada), “The main criterion for pharmaceutical inspection of pills and packaging is the reliability of the equipment. When inspecting pills, you need 100 percent inspection, and a lot of traceability in your system in case there is a failure. That’s the biggest requirement of the pharmaceutical industry,” Perelli said.

Traceability and reliability in a machine vision system typically require effort on both the software and hardware side of the system. “A lot of the tracing is done in hardware, because you can’t depend on software all the time, especially if you have to respond in fractions of a milliseconds; you need hardware to handle that,” Perelli explained.

Hardware, implemented through an intelligent frame grabber/image accelerator card, tracks several common and some less-common events. For instance, hardware is often used to track the beginning, the end of trigger signals, as well as double triggers, trigger bounce and other error conditions. An intelligent frame grabber may also track camera acquisition signals, latency between read commands and image data flow, the image data itself to check for a full frame versus partial frames. Systems with extensive tracking also may track other peripherals, assuming compatibility for such operations, including lighting, PLCs, data storage systems and information networks, such as Ethernet, DeviceNet, Profibus or other industrial networks.

Machine vision frame grabber and/or image accelerator manufacturers depend on some form of acquisition control unit (ACU) or other stand along clock to provide objective time stamps without depending on the clock inside the microprocessor. An encoder located on the conveyor or other material handling system that carries the parts under inspection can also provide this functionality. Main CPU’s, on the other hand, have to run the operating system, track other peripheral input devices non-germane to the vision system, and operate other software that may be open at the time the vision system is running – all of which can introduce latency to microprocessor I/O operations that exceeds the requirements of a real-time machine vision tracing and reliability solution. The best traceability often requires tight integration between the peripheral and image processor; for instance, not all cameras will have alarm functionality to confirm asynchronous triggers or other benchmark events during image acquisition and transmission to the image processor.

Software, according to Perelli, typically handles the data storage and archival of tracking data, completing CFR211’s 100 percent inspection requirement. Software may also track image-processing benchmarks to verify that an image analysis was conducted properly and that the image data is valid. Perelli made one final comment on redundancy in medical inspection applications. “You need lots of storage and redundant storage too, which is why you see a lot of RAID II SCSI controllers in this type of application. Most applications require that you store the image data as well as the pass/fail data,” Perelli concluded.

High performance medical modalities
After product tracking and inspection, medical imaging systems, such as magnetic resonance imaging (MRI), computed tomography (CT), and digital x-rays, are among the largest application areas for machine vision components. According to Joe Sgro, CEO of Alacron (Nashua, NH), making of high performance co-processor imaging acquisition and processing systems, “In general, [medical imaging applications] have fairly large amounts of data flow. These systems have two demands: (1), large data sets and (2), do a lot of work to enhance image. A typical application requires the system to take a lot of pixels, do all kinds of corrections to the images -- including dynamic range, dark and light corrections – possibly do enhancement and/or rotation and modifications as well as identify [regions of interest]. Then take that data for reconstruction [such as combining multiple 2D slices into a 3D image]. Customers tend to do this in one of two ways: with accelerated frame grabbers or clustered computing. Other requirements include interfacing to their sensor system through Camera Link, gigabit Ethernet, or other proprietary interfaces, and support for Linux as well as Windows.”

Although Windows is still dominant today, Sgro added that interest in Linux is growing as more companies go to computer clusters for medical imaging processing and enhancement. Although clusters can offer more image processing horsepower, available space can become a problem. “We have an 8 processor board that’s equal to 8 computers. If you’re not using Blade servers, you might have a space problem,” Sgro advised.

Although eight processors may seem like overkill for most machine vision applications, the processing hungry medical imaging market will take that and more. “Basically, the medical imaging market can go from several gigaflops, which is an accelerated workstation, on up. Things like CT scans can go through the roof, they want as much power as they can get. The more processing power you have, the finer the resolution of the data you can handle in a given amount of time,” Sgro concluded.