Erosions and ankylosis in patients with sacroiliitis are detectable to a high degree of accuracy on CT images using an artificial intelligence (AI)–based algorithm, according to research presented at the 13th International Congress on Spondyloarthritides.
Lennart Jans, MD, head of clinics in musculoskeletal imaging in the department of radiology at Ghent (Belgium) University Hospital, shared data on the development and validation of the algorithm for automatic detection of erosion and ankylosis on CT images of the sacroiliac (SI) joints.
“Essentially, in terms of statistics, this AI algorithm has 95% sensitivity for picking up erosions in patients with clinical symptoms of sacroiliitis, and if this is further developed as a tool, it could aid detection in people with erosions that would otherwise go undetected and undiagnosed,” Jans said in an interview, stressing that the results were still preliminary.
“We want to move from reporting one patient at a time to a system that detects and helps to diagnose larger numbers of patients and makes a larger impact on patient outcomes.”
He stressed that, with thousands of images per patient, it is an impossible workload for any radiology department to read every image necessary to inform diagnoses, and this is only exacerbated by the shortage of rheumatologists, especially in the United States.
Denis Poddubnyy, MD, head of rheumatology at Charité University Hospital, Berlin, acknowledged that AI has potential to improve the recognition of changes indicative of spondyloarthritis (SpA) on imaging. “A standardized, valid, and reliable detection of those changes is relevant for both diagnosis, including differential diagnosis, and classification of SpA.”
Poddubnyy added that the AI-based algorithm developed by Jans and associates is designed to detect very specific SpA structural changes in the SI joints on CT. “CT is usually applied in the clinical practice after MRI … normally in cases where MRI does not provide conclusive results,” he said. Since MRI scans have also been recently used to develop an AI-based algorithm for the detection of active inflammation – not captured by CT – and structural changes in SI joints, he noted that the “generated data on CT should be, therefore, seen in a broader context toward standardization of imaging findings detection.”
Proof-of-Concept Findings Are Due for Scale-Up
Jans acknowledged that the current data only establish proof of concept. Among the study’s 145 patients, 60% were used for training the AI algorithm and 40% for testing it. All patients who had clinical symptoms of sacroiliitis and had undergone a SI joint CT scan were included from two hospitals: Ghent University Hospital and the University of Alberta Hospital, Edmonton. The majority of patients were female (81 of 145). They had a mean age of 40 years, 84 had diagnosed axial SpA, 15 had mechanical back pain, and 46 did not have a final diagnosis.
CT images were examined by three independent and blinded radiologists who annotated erosions more than 1 mm and ankylosis more than 2 mm, while a type of AI algorithm known as a neural network pipeline was developed to segment the SI joints and detect structural lesions.
In the first instance, Jans explained, examination of CT images using the AI algorithm from patients who enter the hospital for other reasons, such as trauma, rheumatic diseases, kidney stones, or appendicitis, might lead to the detection of otherwise unknown erosions. “Often patients have complained of backache for years, seeing various physiotherapists and similar, but had no idea what might be causing it,” he said. “We just don’t have the time for examining all the thousands of images separately. We need some kind of aid here. We need an extra pair of eyes. This is what AI software does.”
Jans said rheumatologists who ultimately want to detect and diagnose patients with SI erosions want to reduce the false-negative findings. “They want the system to pick up all the patients who have erosions. Here, the most important parameter is sensitivity, and we find that our algorithm shows a very high sensitivity. Optimization of the AI algorithm to reduce false negatives resulted in a sensitivity of 95% for detection of erosions on CT of the sacroiliac joints on a patient level.”
While overall accuracy was over 90%, Jans acknowledged that the algorithm was run in a relatively select population of dedicated CT scans of the joints. He is also aware that a good AI algorithm needs to work well across locations and populations. “If you make something within your institution alone, it will not work in a hospital on the other side of the street.”
However, he added, the researchers used images from four different CT scanners and images from two different institutions – one in Canada and their own in Belgium, providing a case mix that makes their algorithm more refined.
Next Step: Test in an Unselected Population
When asked to comment on the study, Mikael Boesen, MD, PhD, of Bispebjerg and Frederiksberg Hospital, Copenhagen, congratulated Jans on the work and remarked that he found the research potentially clinically useful.
“The next steps would be to test the performance of the model in an unselected population of patients who have CT scans of the abdomen for other reasons to test the model’s ability to flag potential SI joint disease to the reader, which is often overlooked, as well as [to see] how the model performs in larger datasets from other hospitals, vendors, and CT-reconstruction algorithms.”
Finally, Boesen pointed out that it would be interesting to see if the AI algorithm can detect different reasons for erosions. “Especially [for] separation between mechanical and inflammatory courses. This could potentially be done by automatically mapping the location of the erosions in the SI joints.”
Jans has now opened up the project to other radiologists to collaborate and provide images to train and test the algorithm further. “We now have 2.4 million images that have been enriched, and we will use these in the near future as we move beyond the proof-of-concept stage.
He is looking for as for as many partners as possible to help collect enriched images and develop this into a real tool for use in hospitals worldwide on clinical patients. “We have joined forces with several hospitals but continue looking for further collaborations.
“We need, just like self-driving cars, not just thousands, but tens of thousands or millions of images to develop this.”
Jans declared receiving speaker fees from UCB, AbbVie, Lilly, and Novartis, and that he is cofounder of a future spin-off of Ghent University RheumaFinder. Poddubnyy and Boesen declared no relevant disclosures.
This article originally appeared on MDedge.com, part of the Medscape Professional Network.