Abstract

Objective: The purpose of the study was to create an Artificial Neural Network (ANN) based on X-ray images of the pelvis, as an additional tool to automate and improve the diagnosis of coxarthrosis. The research is focused on joint space narrowing, which is a radiological symptom showing the thinning of the articular cartilage layer, which is translucent to X-rays. It is the first and the most important of the radiological signs of degenerative changes. Material and Methods: As part of the study, 13374 pelvis cases with 26748 X-ray images of the hip joints were collected. All images were cropped and added to the database with the associated annotations created by the team of three orthopaedists. For the test dataset, 20% of random cases were chosen to correspond to statistical degenerative changes types distribution. The classification task was performed in a two-stage process using Convolutional Neural Networks (CNNs). First, the localization model was trained to locate the bounding boxes (width, height, center coordinates) containing the hip joint with its immediate surroundings, which reduced the size of data and hence the computational power needed for classification. Then, cropped images were classified using another CNN, loosely based on CheXNet architecture. Results: The accuracy of the localization model, measured by the intersection over union metric, was more than 94%. Trained ANN correctly classified 87.4% of cases with a 95% confidence interval (95CI) equals 85.6-89.1%. Results of precision of 91.7% (95CI 90.6-92.7%), sensitivity of 93.5% (95CI 92.0-94.8%), and F1 score of 92.6% (95CI 91.5-93.6%) were achieved. Conclusions: Created ANN reached promising accuracy. Combined with the good automated detection of the hip joints, it could potentially aid the fast differentiation of stable and urgently required medical intervention patients. Acknowledgments: We acknowledge partial support from UE within the NCBiR Epionier program (WG-POPC.03.03.00-00-0008/16-00).

Citations

Authors (8)