Deep Learning Image Reconstruction – what AI looks like in clinical routine

Artificial intelligence is no longer a dream of the future in medicine. Many studies and initial application examples show that it sometimes achieves better results than human physicians. At Jena University Hospital, the work with AI is already lived practice. It is the first institution in the world to use algorithms in radiological routine to reconstruct CT images.

In this MEDICA-tradefair.com interview, Felix Güttler explains why physicians today urgently need the support of intelligent systems, what the use of AI already looks like in everyday clinical practice and what it could look like in the future.

Mr. Güttler, to what extent is there a need for improvement in everyday clinical life, especially in the field of imaging diagnostics?

Felix Güttler: The rapid development and improvement of the various device technologies is increasing the significance and significance of radiological examinations for clinical diagnostics. As a result, we are experiencing an unbroken upward trend in the number of examinations and an increase in image data per patient. In order to counter this growing flood of information, we need intelligent systems that facilitate the analysis of images and keep the workload to a level that doctors can handle.

In order to cope with the increasing workload, you have recently started using Artificial Intelligence in radiological routine. What distinguishes it?

Güttler: We use AI in several places at Jena University Hospital, recently for image reconstruction in computer tomography (CT). Since April of this year, we have been the first institution worldwide to use Deep Learning Image Reconstruction (DLIR) in the broad clinical routine. This enables us to significantly reduce image noise in CT data and thus achieve significantly improved image quality. The underlying AI is based on a Deep Neural Network (DNN). As a rule, AI systems that are approved as medical devices do not learn any more during their use. This is helpful to ensure correct and reproducible behavior. An algorithm is generated from the trained DNN. The neurons of an AI are trained for a certain field of application and then "frozen" for use.

You are using a technology by GE Healthcare. What is so special about it?

Güttler: In my opinion, GE is currently the leader in the field of DLIR for CT. In April 2019, GE was the first company to present a solution to the market that had been approved as a medical device. What was special at that time, however, was not only the uniqueness, but also the form of the AI training at GE. This was done not only on the basis of data already processed using iterative methods, but also on the basis of image information that was not subsequently changed and thus fundamentally more accurate. The confidence in the technical approach and in the generated images was therefore very high right from the start. This trust was very important when establishing a fundamentally new process.

What experience have you gained with this technology in recent months?

Güttler: It quickly became clear to us that DLIR would enable us to face a worldwide upheaval in CT reconstruction procedures. Roughly speaking, we can say that we can measure about 50 percent less image noise than the currently widespread iterative reconstruction methods. In addition, the speed of image reconstruction also improves considerably. We also see great potential for reducing the dose in the future. Savings of between 30 and 80 percent have already been achieved in individual examinations.

We are really impressed so far and hardly use iterative methods anymore. We see positive effects in many questions. For example, DLIR significantly improved contrast resolution in an immunocompromised patient with suspected atypical pneumonia even without a lung filter. DLIR reduced motion artifacts in patients suspected of having suffered a stroke so that a more reliable assessment of the region was possible. These are just a few examples.

Does the AI do the job better than a human being? Which tasks still must be done by human radiologists?

Güttler: In this case, the AI does not perform any medical work. In this respect, the tasks of the radiologist will continue to be completely solved by radiologists. However, they can now access better image data and possibly arrive at a diagnosis faster and more reliably. We are still very far away from generalistic AI systems, as we perhaps know them from science fiction films. At the moment, we have highly specialized applications at AI that serve as a tool for solving a problem.

As in other industries, AI has the potential to automate certain processes and this is accompanied by a change in the professional world. Nevertheless, I am convinced that AI will not displace physicians, for example in radiology. Instead, AI will improve their work reality by, for example, detecting deviations more quickly, automating and prioritizing patient concerns or possible clinical pictures, or performing certain routine tasks completely by the AI. To put it quite banally, where patient welfare, a decision on diagnosis or treatment is concerned, people will always be in control.

What are your other plans regarding the implementation of AI in clinical practice?

Güttler: We do not only see advantages in "pixel AI", as in image optimization. The application of AI in the management of workflows also has special potential. It can be used, for example, to optimize waiting times, optimally utilize devices or automate scheduling. In particular, we are currently working on establishing a manufacturer-independent platform for AI systems for our radiology in order not to be trapped in a particular system landscape, but to work with a partner that also makes it possible to integrate third-party products, for example from start-ups.

However, the use of AI in general offers many more possibilities for medicine. Think, for example, of intelligent control of bed utilization in a hospital or automatic comparison of certain clinical pictures with hundreds or thousands of data points. Imagine you could, for example, fully automate scheduling or have the AI take over routine control measurements. The bottom line for the patient is: simplified processes, faster treatment tailored to individual needs and more time with the attending physician.