Artificial intelligence (AI) technology spans all fields of imaging, including radiation protection, and that has a lot to do with improving image optimisation. But how much can dose be reduced without compromising image quality? Quite a lot, probably, but not too much and only very carefully, experts will explain during a dedicated EuroSafe Imaging session.
When it comes to AI, radiation protection really boils down to dose optimisation. Different methodologies have been developed to confront this challenge, especially in CT, a topic that will be tackled by Mika Kortesniemi, PhD, Chief Physicist and Adjunct Professor at the HUS Medical Imaging Center, University of Helsinki and Helsinki University Hospital, Finland.
“CT is the modality that provides the most cumulative dose to patients. One of the main developments here has been on improving image quality, as this may have the indirect effect of allowing the reduction of dose exposures during the examination, without limiting the diagnostic quality and reliability of the images,” he said.
Deep learning (DL) and convolutional neural networks (CNN) are currently very popular methods among researchers. There has been a very steep rise in the number of scientific results related to DL methods, in dose reduction and other tasks, like diagnostic detection and localisation of morphological tissues. That is because DL is so flexible that it can be used in a wide number of applications in radiology and beyond, Kortesniemi explained.
“DL is like a Swiss army knife. It has many implementations and dose reduction is one of them. CNN and DL can be applied in different areas of radiology, healthcare, and many other sectors, such as the car industry,” he said.
To improve image quality, DL can work from raw data images, reconstructed image data, or both. DL involves the use of neural networks with multiple layers, involving convolutions, non-linear operations, resampling and interconnections in the network, all of which grant DL multiparametric functions. “It means that these layers can basically mimic or approximate any kind of functions or processes we need in image reconstruction,” he said.
DL could help map and convert very low-dose and noisy images into high-quality ones, without increasing dose, and help correct for artefacts. DL methods could also help maintain the images’ natural texture better than iterative reconstruction, where image texture sometimes changes and causes a plastic or watercolour-like appearance.
This potential makes DL-aided image optimisation very interesting for radiologists to work with to lower dose, according to Kortesniemi, a medical physicist. “With DL, we preserve the texture of traditional reconstruction, which means it could be an easier approach for radiology,” he said.
Researchers at Kortesniemi’s institution are looking into clinical validation of these tools within a couple of years. “We are still in the implementation phase with the manufacturer. We have to negotiate a few things, notably regarding the calculation of units, because it demands extra hardware, which is not included in the normal setup. And we need many patients to compare DL reconstruction with traditional and iterative reconstruction, and grade image quality subjectively,” he said.
Researchers also use phantom comparisons, with anthropomorphic phantoms, which are more objective and quantitative measures, to test the visibility of low-contrast targets and help bring early results.
Things are moving fast, and the prospects are engaging.
“First results show that we will get a clear benefit with initial DL techniques, but they still need further evaluation before we can say there is a definite dose reduction with DL in various clinical scan protocols. Thus, you do not just plug and play DL applications; you need to validate them thoroughly,” he concluded.
One can also use AI to improve acquisition parameters, not just image reconstruction. There is not much published work on that topic yet, but Prof. Christoph Hoeschen, from the Institute of Medical Technology at the Otto-von-Guericke University Magdeburg, Germany, will present the information he and his team have collected over the past few months.
“A major question in the development of AI for dose reduction is how much information is really needed from the image itself, and if we can use AI to predict how much dose we need,” he said.
Hoeschen will also talk about a project he is currently undertaking with his team, in which they have committed to reducing dose by two thirds for the same image quality, with CT and more modalities, focusing on interventional imaging. Hoeschen will show how reducing dose further can potentially impair image quality and how this can potentially be tested.
“Some people have claimed they can maintain the same quality with very little dose, by going down to 10%, or even less, of the current dose. With AI, the image looks nice, but you cannot separate what is real information from information that has been generated from the computer without any real reasoning or structure behind it. That is a real problem,” he said.
It is also an issue to use training and testing that are somehow related to each other, he warned, because the study will show results that do not reflect clinical practice.
“Various groups have published things like that, and I will show the problem of such approaches and what happens when you do that. We need some limits and tests that verify accurate approaches,” he concluded.
EuroSafe Imaging Session
EU 18 Artificial intelligence for dose optimisation
- Technology using AI for radiation protection
Mika Kortesniemi; Helsinki/FI
- What is the limit of dose reduction by artificial intelligence methods: 2D and 3D
Christoph Hoeschen; Magdeburg/DE
- Chances and limitations of AI for nuclear medical imaging
Christoph Hoeschen; Magdeburg/DE
Meineke A, Rubbert C, Sawicki LM et al (2019) Potential of a machine-learning model for dose optimization in CT quality assurance. Eur Radiol. 29(7):3705-3713: european-radiology.org/6013
Higaki T, Nishimaru E, Nakamura Y et al (2018) Radiation dose reduction in CT using Deep Learning based Reconstruction (DLR): A phantom study. ECR 2018 / C-1656: myESR.org/181656
Kim DH, Wit H, Thurston MDV (2018) Artificial intelligence in nuclear medicine: automated interpretation of Ioflupane-123 DaTScan for Parkinson’s disease using deep learning. ECR 2018 / C-0668: myESR.org/18668