Cyth Integrates Neural Vision for Medical Device Inspection
Introducing automation into your product inspection process has many benefits to your business, such as increasing efficiency and decreasing costs. However, in the medical device industry, taking the necessary steps to ensure that each product is inspected to a high medical standard can end up preventing complications in surgeries or other medical operations performed. A global medical device manufacturer needed an automated vision system that was designed to load their product, run it through an entire inspection process, and identify if the product was up to industry standard or if it was defective. Their product is a guide wire that is used in surgery or other medical settings where doctors will feed it through a vein or artery to remove clots or other similar procedures. If the wire has any sharp edges or was cut improperly, it could be lodged inside the body or cause other harm. The customer reached out to Cyth for help developing a custom application that could inspect these wires and ensure the client was manufacturing the highest quality devices possible.
Client Request & Cyth’s Solution The product needing inspection is a long metallic wire with a tip that has two sets of finely wound coils. There are three different solder joints connecting the guide wire. One solder joint to connect the wire to the coils, one joining the two coils together, and one as a bead to cover up where the coil is clipped to cover any sharp edges. The guide wire is made up of miscellaneous metal compounds and is approximately two feet in length. The inspection focuses on the 0.3-millimeter width of two inches of the coil.
The client provided a mechanical system that moves and rotates the wire and manages the PLC software, sending out triggers for the cameras to take photos. However, they had no interface for receiving the captured images and saving them for post-processing. Cyth’s team was challenged to build the GUI and software that would connect with the cameras and the HMI. The vision system consisted of two 5MP cameras from Cognex with two telecentric lenses from Edmund Optics. LabVIEW was used to integrate the hardware interface, and Cyth’s Neural Vision was utilized to examine each captured image and save it along with the corresponding information to a local database. After collecting a substantial number of images, Neural Vision begins analyzing each image through the power of deep learning. The software can determine the length of the coil as well as inspect the shape of the tip to ensure that it was properly formed and soldered. The user interface will show a “pass/fail” result for each wire if the length is too short, if the wire was soldered improperly, or if one of the tip shapes scored poorly.
The second project phase was a “build out” of the shape inspection. In the first phase there was a false reject percentage of 13%. This means 13% of the time, the operator had identified the part as good while the system determined it a failure. The main goal of the second phase was to reduce the false reject percentage to under 10%. To accomplish this goal, the team categorized the different kinds of tip shape defects rather than labeling all defects as a collective “bad” product. Being more specific would tailor down the image results to consistent accuracy. Cyth’s engineers were able to label ten different defect categories with individual thresholds. The frequency of false rejects was reduced to only 6%.
Challenge The largest obstacle to overcome was determining how to accurately program the Neural Vision software by tagging images in the most effective way. It was challenging to decide how to best analyze the tip shape results, and this problem could only be solved through trial-and-error and better educating the client.
Tagging your images is one of the most crucial steps in the development process. To develop a robust and accurate model, the user must correctly tag product images as “good” or “bad”. “Programming” the Neural Vision software by tagging the images takes strong attention to detail when inspecting tiny defects on a relatively large image. If you accidently tag a “bad” image as “good”, the Neural Vision model can become confused as it can’t detect the differences between the two. The operator who is tagging images must be specific and consistent, even over hundreds of images. Each tagged image adds to the model’s ability to identify product defects in a subjective manner, achieving an accurate statistical analysis report.
Outcome By the end of the project, Cyth had met the client’s main goal of reducing the percentage of false rejects and creating a system that could identify the defective wires. The client was given the opportunity to characterize these defects more specifically, rather than with a mere “pass/fail” system. This allowed the system to provide a detailed report, explaining the reason why the product failed and helping the client to identify errors that might have occurred in the production process. Cyth’s engineering team collaborated with the client to provide as much customer education as possible, giving them complete control of their vision inspection system performance.
Transitioning an inspection system to be automated doesn’t only save labor, time and money as it increases efficiency, but it also increases the consistency of the product quality. Human error is decreased as the system relies on machine vision technology, assuring that each wire that passes through the inspection will be up to the standard necessary for medical usage. The client now has a functioning inspection system that will allow them to examine more devices than ever before, and they can trust the devices they sell will always be of the highest quality. Technical Specifications • 2, GigE, 5MP, 24fps, CMOS, Sony IMX264 2/3", Monochrome Cameras • 2X, 65mm WD Compact Telecentric Len • 52mm Telecentric Backlight Illuminator • Blue, 0.315" LED Spot Light • Cyth’s Neural Vision Software • LabVIEW 2017 (64-bit)