Neural network learns to differentiate between healthy and inflamed bones using finger joints.
Researchers have been able to teach artificial intelligence neural networks to distinguish between two different kinds of arthritis and healthy joints. The neural network was able to detect 82% of the healthy joints and 75% of cases of rheumatoid arthritis. When combined with the expertise of a doctor, it could lead to much more accurate diagnoses. Researchers are planning to investigate this approach further in another project.
This breakthrough by a team of doctors and computer scientists has been published in the journal Frontiers in Medicine.
There are many different varieties of arthritis, and determining which type of inflammatory illness is affecting a patient’s joints may be difficult. Computer scientists and physicians from Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) and Universitätsklinikum Erlangen have now taught artificial neural networks to distinguish between rheumatoid arthritis, psoriatic arthritis, and healthy joints in an interdisciplinary research effort.
Within the scope of the BMBF-funded project “Molecular characterization of arthritis remission (MASCARA),” a team led by Prof. Andreas Maier and Lukas Folle from the Chair of Computer Science 5 (Pattern Recognition) and PD Dr. Arnd Kleyer and Prof. Dr. Georg Schett from the Department of Medicine 3 at Universitätsklinikum Erlangen was tasked with investigating the following questions: Can artificial intelligence (AI) recognize different forms of arthritis based on joint shape patterns? Is this strategy useful for making more precise diagnoses of undifferentiated arthritis? Is there any part of the joint that should be inspected more carefully during a diagnosis?
Currently, a lack of biomarkers makes correct categorization of the relevant form of arthritis challenging. X-ray pictures used to help diagnosis are also not completely trustworthy since their two-dimensionality is insufficiently precise and leaves room for interpretation. This is in addition to the challenge of placing the joint under examination for X-ray imaging.
Artificial networks learn using finger joints
To f ind the answers to its questions, the research team focused its investigations on the metacarpophalangeal joints of the fingers – regions in the body that are very often affected early on in patients with autoimmune diseases such as rheumatoid arthritis or psoriatic arthritis. A network of artificial neurons was trained using finger scans from high-resolution peripheral quantitative computer tomography (HR-pQCT) with the aim of differentiating between “healthy” joints and those of patients with rheumatoid or psoriatic arthritis.
HR-pQCT was selected as it is currently the best quantitative method of producing three-dimensional images of human bones in the highest resolution. In the case of arthritis, changes in the structure of bones can be very accurately detected, which makes precise classification possible.
Neural networks could make more targeted treatment possible
A total of 932 new HR-pQCT scans from 611 patients were then used to check if the artificial network can actually implement what it had learned: Can it provide a correct assessment of the previously classified finger joints?
The results showed that AI detected 82% of the healthy joints, 75% of the cases of rheumatoid arthritis, and 68% of the cases of psoriatic arthritis, which is a very high hit probability without any further information. When combined with the expertise of a rheumatologist, it could lead to much more accurate diagnoses. In addition, when presented with cases of undifferentiated arthritis, the network was able to classify them correctly.
“We are very satisfied with the results of the study as they show that artificial intelligence can help us to classify arthritis more easily, which could lead to quicker and more targeted treatment for patients. However, we are aware of the fact that there are other categories that need to be fed into the network. We are also planning to transfer the AI method to other imaging methods such as ultrasound or MRI, which are more readily available,” explains Lukas Folle.
Hotspots could lead to faster diagnoses
Whereas the research team was able to use high-resolution computer tomography, this type of imaging is only rarely available to physicians under normal circumstances because of restraints in terms of space and costs. However, these new findings are still useful as the neural network detected certain areas of the joints that provide the most information about a specific type of arthritis which is known as intra-articular hotspots. “In the future, this could mean that physicians could use these areas as another piece in the diagnostic puzzle to confirm suspected cases,” explains Dr. Kleyer. This would save time and effort during the diagnosis and is already in fact possible using ultrasound, for example. Kleyer and Maier are planning to investigate this approach further in another project with their research groups.
Reference: “Deep Learning-Based Classification of Inflammatory Arthritis by Identification of Joint Shape Patterns—How Neural Networks Can Tell Us Where to ‘Deep Dive’ Clinically” by Lukas Folle, David Simon, Koray Tascilar, Gerhard Krönke, Anna-Maria Liphardt, Andreas Maier, Georg Schett and Arnd Kleyer, 10 March 2022, Frontiers in Medicine.