Analysis and development of Convolutional Neural Networks for diagnosis of Osteoarthritis in knee radiographs of the Brazilian Longitudinal Study of Adult Health Musculoskeletal (ELSA-Brasil MSK)

Abstract

Introduction: Knee Osteoarthritis (KOA) is frequently diagnosed by evaluating radiographies according to the Kellgren-Lawrence (KL) grading system, which involves the assessment of osteophytes and articular joint-space narrowing. In large-scale longitudinal epidemiological studies, such classification is performed by specialist physicians, in a laborious and time-consuming process that requires rigorous training. In this context, the application of computational models for automated KOA classification can contribute to the diagnostic flow, reducing the total number of exams to be evaluated by humans. Among the automatic classification models, convolutional neural networks (CNN) stand out, having demonstrated promising results in the medical imaging diagnosis field. Recent studies have shown, however, that the development and validation of currently existing CNN is concentrated in the population of developed countries, demonstrating the need for greater external validation of such methods. Objectives: To evaluate the performance of a previously published CNN for KOA diagnosis on radiographs of ELSA-Brasil MSK participants and to develop a CNN from scratch trained with the ELSA-Brasil MSK baseline exams. Method: This is a cross-sectional study carried out with radiographies from ELSA-Brasil MSK’s baseline. The ELSA-Brasil MSK, an ancillary study to the Longitudinal Study of Adult Health (ELSA-Brasil), included, in its baseline, 2901 active or retired public servants from two large teaching and research institutions in Minas Gerais. To verify the performance of the previously published CNN, a convenience sample consisting of 243 radiographs was selected. CNN predictions were compared with the KL classification and OA diagnosis performed by ELSA-Brasil MSK’s radiologists. For the development of the CNN, all radiographies from the baseline of the ELSA-Brasil MSK (5660 knees) were used. The exams were interpreted by a radiologist with specific training and calibration, previously published. Results: The previously published CNN presented an area under receptor operating characteristics curve (AUC) of 0.901 for the diagnosis of OA (CI 95%: 0.858-0.945). The CNN developed from the radiographs and classifications of the ELSA-Brasil MSK presented an (AUC) of 0.866 (95% CI: 0.842-0.882). Conclusions: The evaluated CNN presented similar performance for the evaluation of the ELSA-Brasil MSK radiographies. The use of such models can help as a screening tool, reducing the total number of exams evaluated by the radiologists of the study, and/or as a dual reading tool, contributing to the reduction of possible interpretation errors.

Ähnliche Arbeiten