Editor’s note: EULAR 2020, the annual European Congress of Rheumatology, which was originally scheduled to be held in Frankfurt, Germany, starting June 3, was moved to a virtual format due to the COVID-19 pandemic.
EULAR 2020 e-CONGRESS—Medicine is often at the forefront of technological revolutions in society. Indeed, the accomplishments of clinicians and researchers can demonstrate what’s possible when theory is translated into practice—from bench to bedside. One area currently of great interest is how artificial intelligence may be able to assist physicians with patient care. During the 2020 European e-Congress of Rheumatology, June 3–6, a fascinating session, titled Artificial Intelligence and Machine Learning in Rheumatology Imaging: Are We Ready?, explored this topic in great detail and discussed work being conducted around the world.
The session started with a synopsis of the history of, and description of terms that apply to, artificial intelligence and machine learning. Artificial intelligence is the ability of a computer to perform tasks that generally require the intellect of a human being. Machine learning is an application of artificial intelligence focused on how computers learn from data and involves the intersection of statistics and computer science through the use of efficient computing algorithms.1 Neural networks are a form of machine learning.
In 1943, the neuroscientist Warren McCulloch and the logician Walter Pitts proposed the first computational model of a neuron. Each McCulloch-Pitt neuron had binary inputs and binary outputs (i.e., 0 or 1). Inputs could be excitatory (i.e., positive) or inhibitory (i.e., negative). The sum of the inputs could then be calculated for a connected chain of neurons. If the final sum was above a pre-specified threshold, then the output was 1. If the sum was below that threshold, it was 0.2 From this model, neural networks have continued to evolve and become more complex, but always with the foundational concept that neural networks are a series of algorithms that identify relationships in a data set through a process that mimics how the human brain functions.3
Imaging-Based Scoring
Several groups that have applied machine learning techniques to problems facing rheumatologists presented their findings in the EULAR session. Thomas Deimel, MD, Division of Rheumatology, Medical University of Vienna, Austria, offered a presentation on autoscoRA, a deep learning model that enables the automatic scoring of the radiographic progression of rheumatoid arthritis (RA).
Standardized scoring systems of radiographs, such as the Sharp/van der Heijde score, are time intensive, require specially trained staff and have a fair amount of intra-reader variation. Dr. Deimel and colleagues used a data set of more than 5,000 hand radiographs from 640 adult RA patients to train, validate and test the automated, deep-learning computer system.4 Using the human-rated Sharp/van der Heijde score as the ground truth, only 1.8% of metacarpal phalangeal (MCP) joints and 1.7% of proximal interphalangeal (PIP) joints scores calculated by the deep learning system differed from the Sharp/van der Heijde score by more than one point. The researchers note this is the first work to automate manual scoring on a large scale. Future work will further refine the system’s accuracy.
Anders Christensen, MSc, Health and Welfare Technology, University of Southern Denmark, Odense, presented work on an automated scoring system for disease activity in RA patients using analysis of ultrasound images. The algorithm developed and trained by this group is a “convolutional neural network,” an arrangement of many layers of smart filters that, when correctly assembled and trained, can interpret and analyze complex data, such as medical images. The creation of a convolutional neural network is a supervised learning process, and the trainers of the computer are the rheumatologists who provide the ground truth (i.e., the correct outputs that the algorithm is supposed to produce). Mr. Christensen and colleagues note their system is a cascade of convolutional neural networks, which allows it to follow a hierarchy of decisions about the grade of synovial hypertrophy and synovitis to determine the correct, ultimate conclusion.
In evaluating more than 1,600 ultrasound images, the model achieved 84% accuracy. On a per patient level, no significant difference existed between the classifications of the computer model and the human experts. The research team indicated its data show convolutional neural networks with the cascade architecture can be used by rheumatologists to identify disease activity in RA patients.
In another presentation, Alarico Ariani, MD, from the Unit of Internal Medicine and Rheumatology, Azienda Ospedaliero Universitaria, Parma, Italy, described the use of a computer-assisted, automated quantitative computed tomography (QCT) system to evaluate interstitial lung disease in systemic sclerosis. In 2017, Dr. Ariani and colleagues showed the QCT method could discriminate between well-defined mortality risk categories, as well as clinical prediction models.5
Dr. Ariani aired a video showing how all needed clinical data about a patient could easily be entered into the QCT system to identify patients with a high risk of mortality within 10 years—all in less than one minute. This type of work in predicting mortality risk will play an essential role in determining which patients may benefit from new treatments, such as anti-fibrotic therapy.
More Advances for Rheumatology
The remainder of the presentations were equally compelling, describing the use of artificial intelligence to help with challenges in rheumatology. One group demonstrated the predictive value of bone texture features, extracted by deep-learning models, in the detection of osteoarthritis. Another research team described detection of subclinical skin manifestations in patients with psoriasis and psoriatic arthritis using fluorescence optical imaging, a technique that can evaluate changes in the microvasculature of the hands.
These advancements illustrate how rheumatologists, radiologists, computer scientists and others will likely collaborate in the future to make faster and more accurate diagnoses. The future appears bright for the use of artificial intelligence in rheumatology imaging, and this future may be fast approaching. It does not require a great deal of imagination to see how artificial intelligence can buttress human intelligence and help doctors and patients in our quest to think smarter and faster.
Jason Liebowitz, MD, completed his fellowship in rheumatology at Johns Hopkins University, Baltimore, where he also earned his medical degree. He is currently in practice with Skylands Medical Group, N.J.
References
- Deo RC. Machine learning in medicine. Circulation. 2015 Nov 17;132(20):1920–1930.
- Sinha U. First artificial neurons: The McCulloch-Pitts model. AI Shack. 2020.
- Chen J. Neural network. Investopedia. 2020 May 20.
- van der Heijde D. How to read radiographs according to the Sharp/van der Heijde method. J Rheumatol. 2000 Jan;27(1):261–263.
- Ariani A, Silva M, Seletti V, et al. Quantitative chest computed tomography is associated with two prediction models of mortality in interstitial lung disease related to systemic sclerosis. Rheumatology (Oxford). 2017 Jun 1;56(6):922–927.
