Schulenburg (VisiConsult): Setting up a scan is an optimization problem, and one that machines are really good at solving, while humans are not. That doesn’t necessarily always require AI. It might help, but it doesn’t have to. In any case, algorithms are ideal for this task. So the answer is yes. The use of algorithms for this task offers many possibilities and high potential.
Mann (Hexagon): Another thing is the optimization of automated inspection planning, which have already been carried out previously. What we learned with ChatGPT is that generative AIs are emerging and making the know-how available. If we transfer the know-how that we have into a database where someone can search for it and e.g. a consultants who wants to measure an airplane will only takes three points, then the system will give him advice and say, okay, you’ll get a perfect plan but you should take at least five points. In this case, all other operators would deduct around ten to 20 points. AI will take care of that, but it will provide a lot of support to the operator, so they may be able to operate different systems without being an expert on each one.
Schulenburg (VisiConsult): We have a little wizard tool that alerts the operator that maybe the kV is a little too low, the mA needs to be a little higher. There are scores that give hints that you don’t have penetration, you could improve your resolution, you could improve your contrast and so on. This is a little tool, and it’s not necessarily always AI, but really helps operators to do a better job.
Can AI improve inline 3D metrology because this is an area where you want to have a good raw data quality?
Mann (Hexagon): Using AI to improve data collection can be helpful. We are not striving to optimize precision to the maximum, but to optimize the raw data so that it is more stable and performant and then you can scan it faster and get a better image, even in inline applications.
Schulenburg (VisiConsult): I studied IT and image processing a long time ago and the principle I have learned back then is: Garbage in, garbage out. Having better quality in your raw data, significantly boosts the results. Applying better algorithms for detection and analytical tasks gives you another advantage. However, the first task should always be to improve the physical parameters of a scan, use the right kV, the right mA, the right filter and settings, then improve the reconstruction, improve the data going into the analysis and then concentrate on the analysis. At every stage of this chain AI has potential.
Wojek (Zeiss): This basically confirms what we do for different cases. We build AI into the reconstruction workflow and then include it in your evaluation until the end of the workflow to achieve the best results.
What role do synthetic images play in the use of AI?
Schulenburg (VisiConsult): They are useful when you don’t have enough data, but they can also be dangerous because they can lead to worse model performance as there typically is a bias in the synthetic data and nothing replaces real data. It’s a great proxy to get started if you have no or very few real data. It also can enrich a data set that contains rare defects. Synthetic data should always be replaced with real data over time as you generate more data, to improve the model performance over time. We have developed a generative network, that creates these artificial defects. You just have to be careful how you use it. It is a helpful tool, but also dangerous if used incorrectly.
Wojek (Zeiss): Especially when starting projects, the most important thing is to have a lot of data as a starting point, and usually you don’t have a lot of data from your application at the beginning. Then synthetic generation sometimes helps even as a simple extension. If you understand the process very well, you can even generate a lot of good data through simple image manipulation to get started. You can also use tools such as generative AI, to generate data. These are a little more difficult to handle because the parameters are a little trickier but definitely this is the right way.
Are there CT applications where AI is not desirable for certain reasons?
Schulenburg (VisiConsult): You have to monitor the results very carefully if you use any time of generative AI. For example, an AI based reconstruction in CT. What could eventually happen is that the AI reconstruction gives you what you want to see, but not what really exists. Noise reduction may allow you to get a sharp image but at the same time eliminate indications and imperfections. Both of these tools can work but need to be carefully verified to see if it works in your application. If it’s an image processing step it really should only be applied when it’s useful, and it should be applied in a good and supervised way, and against a known ground truth. Anything else can be potentially dangerous.