Use in Imaging
Dr. Hügle lists a number of benefits and examples of AI and machine learning in imaging for rheumatology. AI can:
- Assist radiologists and rheumatologists in improving efficient, accurate diagnosis. Example: AI analysis of musculoskeletal diseases is relatively easy given the binary questions asked (e.g., Are osteophytes present or not? Is joint space narrowing present or not? Is there calcification in the meniscus or not?). This is the type of repetitive task that provides an inroad to AI and machine learning;
- Generate automated reports, reducing the time clinicians spend on dictating radiologic reports, such as dual-energy X-ray absorptiometry findings, and providing more time for difficult cases or interacting with patients;
- Reduce the number of images needed for patients by automatically detecting and identifying features in images taken for other purposes. Example: An abdominal CT scan performed for other reasons may automatically provide data on bone density; and
- Allow faster, better, large-scale scoring of images for clinical trials, for example in osteoarthritis. Automated image analysis will eliminate variability among researchers.
A number of automated radiological image processing software applications, such as one for opportunistic bone density in CT scans, are under review by the U.S. Food & Drug Administration, says Dr. Hügle.
“I think that within two to three years, AI will become a reliable partner for rheumatologists in supporting image-based diagnosis,” Dr. Hügle says, adding that he thinks AI will “sneak in” to the daily lives of rheumatologists, and once it becomes familiar it will quickly become essential and normal.
Glossary of Terms3,6
| Artificial Intelligence | Field of computer science since the 1950s, referring to the creation of computer systems to perform a task that typically requires human intelligence. |
| Machine Learning | A subfield of AI since the 1980s in which a computer learns automatically from the data presented to it, mostly in a supervised learning situation. |
| Deep Learning | A subset of machine learning since the 2010s that uses artificial neural networks with multiple layers to decode (imaging) raw data. |
Learning Curve
To help rheumatologists understand how AI and machine learning may be applied to imaging in rheumatology, Dr. Stoel outlines different ways these tools may be used to interpret digital images:3
- Fully human interpretation (mostly used in daily clinical practice): Images are interpreted by a clinician by first detecting abnormal structures in the image (i.e., detection) and subsequently determining what type of lesions they are (i.e., classification). In this approach, AI is used only to produce the images by reconstructing images from raw information from the scanner and by enhancing the appearance of the images;
- Hybrid or computer-aided approach: The computer helps the clinician detect suspected areas and/or classify lesions. Because computers (and in some cases clinicians) always need numerical measurements to perform classification, certain characteristics of the lesion need to be measured first. This intermediate step is called quantification. Usually these characteristic features are specified by a computer scientist in consultation with a clinician (e.g., shape, size, brightness; also known as hand-crafted features); and
- Fully automatic interpretation: here the whole process is automated fully. This can be done by still following the consecutive steps of the hybrid approach—automatic detection, quantification and classification—but hand-crafted features would still need to be specified by human researchers. This could be avoided by letting the computer learn discriminative features independently, without human interaction, by using machine-learning techniques, such as artificial neural networks.
When applying AI to imaging in rheumatology, Dr. Stoel explains that automation may be preferred and is particularly relevant for clinical trials due to reductions in cost, time and observer variability, as well as requiring less training and fewer skills.
He highlights ongoing research on automated image analysis methods to detect and assess the main musculoskeletal components of rheumatoid arthritis (RA)— synovitis, tenosynovitis and bone marrow edema, bone erosions and cartilage loss. Most research has gone into automatic quantification of cartilage loss, with early developments in AI using plain radiographs to quantify joint space in the hand and knee, and more recent AI methods to automatically measure joint spaces within the wrist.
Newer developments focus on using AI methods with magnetic resonance imaging scans to measure cartilage thickness and volume, and deep learning to detect knee cartilage and further classify cartilage lesions.
