A radiomics-based machine-learning model can detect structural damage in hand osteoarthritis from routine x-rays, according to a study published May 21 in Osteoarthritis and Cartilage.
The finding suggests that radiomics may offer a pathway toward objective, reproducible, and fully automated scoring of hand osteoarthritis severity, reducing reliance on subjective visual grading, reported lead author Laetitia Perronne, MD, of Université Paris Cité in France, and colleagues.
"This is the first study applying radiomics to standard hand radiographs for automated and objective scoring of radiographic severity in [hand osteoarthritis]," the authors wrote.
Hand osteoarthritis (HOA) affects approximately 8% of men and 16% of women over the age of 50. The current standard for grading the disease on x-rays is the Kellgren-Lawrence (KL) scoring system, a five-point scale originally developed for knee osteoarthritis. Despite its widespread use, the KL system is susceptible to subjectivity and lacks sensitivity in detecting early and moderate disease, according to the authors.
Radiomics, which extracts quantitative intensity, shape, and texture features from images, could offer a more accurate approach, the authors hypothesized. To that end, the group collected hand x-rays from 374 adults (83% women) with a clinical diagnosis of HOA who presented at their center between January 2017 and December 2020.
The researchers used a U-Net model to semiautomatically segment the x-rays and extract radiomics features describing intensity, shape, and texture for each joint. Next, they trained random forest classifiers on three tasks: detecting any structural involvement (KL ≥ 2), detecting severe disease (KL 3–4), and predicting the full multiclass KL grade distribution (KL 0–4), across a total of 10,395 analyzed joints.
Overlap between manual (radiologist) and semiautomatic (algorithm) segmentations of hand joints on a standard posteroanterior radiograph. Orange overlay indicates manual radiologist segmentations, blue overlay shows algorithm segmentations, and green overlay indicates consistent regions between both methods.Osteoarthritis and Cartilage
“This study demonstrates the feasibility and potential of applying radiomics to standard hand radiographs for semi-automated assessment of structural severity in HOA,” the group wrote.
The researchers noted limitations, namely that the study relied on x-rays scored by a single musculoskeletal radiologist, which ensured internal consistency but precluded formal interobserver reliability analysis. In addition, the single-center design and predominantly female cohort may limit the study’s generalizability, they wrote.
“External validation, integration of multimodal data, and the development of anatomically informed or deep learning-based models remain essential for clinical translation,” the researchers concluded.
The full study is available here.
