Role of AI in Musculoskeletal Ultrasound

Rudra Prosad Goswami MD DM
Associate Professor
Department of Rheumatology, All India Institute of Medical Sciences, New Delhi, India

Your experience of using AI in MSK USG

We from AIIMS, New Delhi collaborated with  Dr Arindam Banerjee, from Department of Mathematics, IIT, Kharagpur. We were interested to see if one of the machine learning (ML) algorithms can classify MSK USG images of the firstt MTP joints for the detection of DCS. It turned out that one of the models, called XGBoost, did quite well. When we compared the performance of the algorithm to three independent and blinded clinicians, the algorithm outperformed the “average” of the three, but not the best performer of the three. The number of tested images were pretty small, just above 400, so we still have a lot more to do.

Current status of AI integration into musculoskeletal ultrasound

It is not routinely integrated but there are papers on using ML algorithms for reading SI joint X rays or hand X ray of CT scans for detection of lung tumors etc. currently the best application is generating images for teaching students. Training and choice of algorithms and post-processing of images are key steps before any of these algorithms can be implemented in routine clinical practice.

Potential benefits of AI-driven ultrasound technologies

An obvious benefit is reducing errors in detection of patterns. But none of the models will detect any abnormality more than what an expert has already detected.

Role of AI algorithms in identifying subtle changes in joint pathology

ML algorithms can assist in identifying “subtle” changes given that they are trained to identify, first the region of interest and then “normal” and then “abnormal”. All these will follow the prerequisite that they are trained on good quality images on representative samples and these “abnormalities” are correctly codified in the images on which the algorithm is trained. It is not expected for an algorithm trained to identify joint space narrowing in SI joints to identify joint erosion in metacarpophalangeal joints.

Challenges and limitations of implementing AI in ultrasound for rheumatology considerations

Major challenge is collecting enough images. A sample size in hundreds, which is generally considered good for most biological studies, is grossly inadequate for AI/ML studies. We need more collaborations. Another point also needs mention is that we observed tree algorithms giving better results than Neural Net indicating that the test and training sets are very similar and from a homogeneous population and finally a small sample, the first of which is an advantage and the second a disadvantage.

 Suggested reading

  1. Lee S, Jeon U, Lee JH, Kang S, Kim H, Lee J, Chung MJ, Cha HS. Artificial intelligence for the detection of sacroiliitis on magnetic resonance imaging in patients with axial spondyloarthritis. Front Immunol 2023;14:1278247. 
  2. Miyama K, Bise R, Ikemura S, Kai K, Kanahori M, Arisumi S, Uchida T, Nakashima Y, Uchida S. Deep learning-based automatic-bone-destruction-evaluation system using contextual information from other joints. Arthritis Res Ther 2022;24(1):227.