POS0891 AUTOMATIC COMPUTATION OF KNEE OSTEOARTHRITIS SEVERITY USING KNEE X-RAYS AND CONVOLUTIONAL NEURAL NETWORKS

Background

Knee osteoarthritis is a heterogeneous and complex degenerative pathology, characterized by a progressive deterioration of bone cartilage and structural modifications of the joint [1]. The precision of the diagnosis and the rating of the severity are major criteria for the therapeutic management and its follow-up. They are based on three criteria: the assessment of the pain, of the functional impairment and of the structural modifications. For this last criterion, the standard protocol in routine care remains the interpretation of X-ray images using standardized scales. The Kellgren-Lawrence (KL) score, which assesses both the joint space and the presence of osteophytes, allows a classification of the stages of osteoarthritis, but it relies on subjective manual interpretation and is time consuming for practitioners [2].

Objectives

In this study, we have developed artificial intelligence algorithms to automatically measure the tibia-femur joint spacing (or joint space width JSW) and determine the Kellgren-Lawrence (KL) score.

Methods

We constituted a retrospective cohort of 19,560 patients. Using all their images, we trained different neural networks in order to select just knee AP X-rays without prosthesis nor artifacts. Our work explores two approaches: the prediction of the stage of osteoarthritis according to the KL scale and the measurement of the JSW. For the prediction of the KL score, 2,081 X-rays annotated by 3 radiologists were used to train a convolutional neural network (CNN). The measurement of the JSW required the realization of 3 different annotations: the positioning of the joint, of the two condyles (medial and lateral) and the contouring of tibia and femur. Three neural networks were optimized to reproduce these annotations before calculating the JSW for each condyle. For each individual task, we decomposed the datasets into training, validation, and test sets, used different data augmentation techniques, and researched the best possible architecture.

Results

The Kellgren-Lawrence score prediction obtained the following performances: an accuracy of 0.92, a sensitivity of 0.84 and an average area under the ROC curve (AuC) of 0.97. To evaluate the measurement of the JSW, we calculated the correlation between the area measured by the annotators and the area predicted by the algorithms, obtaining a Pearson correlation of 0.84.

Conclusion

This study highlights the relevance of the use of artificial neural networks for the assessment of osteoarthritis. Their performance opens the way to a tool assisting in the precise and standardized gradation of the severity of joint degradation.

References

[1]Lawrence, J. S., Bremner, J. M., & Bier, F. (1966). Osteo-arthrosis. Prevalence in the population and relationship between symptoms and x-ray changes. Annals of the rheumatic diseases, 25(1), 1. [2]Kellgren, J. H., & Lawrence, J. (1957). Radiological assessment of osteo-arthrosis. Annals of the rheumatic diseases, 16(4), 494.

Acknowledgements:

NIL.

Disclosure of Interests

None Declared.