Artificial Intelligence (AI) and reconstruction: this is not a movie!
Warning: one train can hide another
If you have not seen the short video that demonstrates the capacities of AI in a spectacular fashion, I would recommend that you follow the link below: https://vimeo.com/455471233. In it, we see how one of the first scenes ever filmed, which dates back to 1895, has been upgraded to 4K super resolution using AI. As you can see for yourself, it is quite spectacular. But, in our business, can we invent information? Let me explain further: look at the lady in the foreground, her face has been reshaped by statistical averaging, but what if, in real life, the lady was one-eyed. No AI could detect that. For the cinema it’s great, but for making a diagnosis it is very dangerous.
We must therefore be wary of “super system” approaches, which erase what we want to see from a diagnostic point of view. There is no problem with that in a film, it just needs to look “pretty” so only the scenery counts; for a diagnosis the scenery is of little importance, only the detail counts.
The “super system” approach consists of thinking: “I have access to a super scanner, a machine with hyper resolution: a prototype that I would never have in the clinic. I will build a database on my super scanner and on my traditional scanner. Then I will reread the images using an AI approach that allows me to over sample my images. Lastly, I will apply this program to the images in my traditional scanner and after the training that corrects and modifies my data, I can obtain images of a quality equivalent to those of the super scanner with my traditional scanner.
With this in mind it is a very efficient tool for 3D reconstruction and imaging
So it’s not actually any use? It is just a case of spitting in the wind? No, that is certainly not the case in our field of 3D volume imaging. When it is correctly understood, this technology will offer major advances in medical diagnosis, in detecting threats from baggage screening systems and in searching for defects in manufactured parts. The technique and many of the associated tools are reliable and effective and relatively easy to implement for specialists in the field of GPU reconstruction and calculation. OF course, there are still technological challenges but these are expressed more in terms of miniaturization and energy consumption, typically for embedded applications. If your application requires a PC, these obstacles do not exist and you can therefore take action very rapidly.
What to do and how to do it
As you have certainly understood, you must first stop believing in miracles: it is revolutionary but not miraculous. It is therefore extremely important to discuss and set achievable goals with your partner in terms of diagnosis, artifact removal, image segmentation, etc.
The main pitfall, since we have seen that the technical aspect and programming are controlled, comes from the knowledge required to develop and refine your model. In fact, a considerable number of AI projects have failed because the databases either do not exist or contain very little information. Our advice is: your results will very often be “application dependent”. Do not start with too broad a theme, focus on a point where your business expertise is identifiable and reproducible. With these guidelines and a solid partner, the results should come rapidly.