Deep learning can be used for accurate classification of hand osteoarthritis using plain radiographs

Hand osteoarthritis poses a growing health concern, necessitating accurate and efficient diagnostic methods. Convolutional neural networks offer a promising approach for automating osteoarthritis classification, but their application to hand osteoarthritis remains underexplored. This study aimed to develop and evaluate a convolutional neural network model for detecting and classifying hand osteoarthritis on plain radiographs according to the Kellgren-Lawrence (KL) and modified Eaton-Littler (EL) classification systems.

A total of 7515 examinations were retrospectively collected (27,965 images) from one hospital (2002-2016) and split into training (n = 7,100), validation (n = 415), and test (n = 327) sets with no patient overlap. Images were labelled for presence, grade, and joint involvement (interphalangeal, metacarpophalangeal, thumb carpometacarpal, and scaphotrapeziotrapezoidal joints). A residual-network model with 39 convolutional layers was trained from scratch in PyTorch with data augmentation (random rotation, flipping, and cropping).

The convolutional neural network model achieved adequate performance in detecting hand OA (AUC = 0.91, sensitivity = 90%, specificity = 83%) and classifying osteoarthritis grades. However, lower osteoarthritis grades were challenging, particularly in the CMC1 and STT joints. Despite limitations in inter-reader reliability of classification systems, the convolutional neural network model demonstrated capability in localising osteoarthritis findings.

This study demonstrates the feasibility of convolutional neural network models in automating hand osteoarthritis detection and classification. The model's adequate performance suggests its potential for clinical use in enhancing classification accuracy and streamlining workflows. Challenges in lower-grade osteoarthritis classification highlight the need for improved standardisation in classification systems. Further research is warranted to validate and refine the model's performance in diverse clinical settings and assess its real-world implications.

Level III.