Recognition of Bone Fractures in X-Ray Images

Ten of thousands of musculoskeletal radiographs are diagnosed every year in Israel. Providing reliable radiograph diagnosis is considered as a major healthcare challenge. Right fracture diagnosis is crucial for appropriate treatment. Wrong diagnosis could lead to wrong procedures or other not required interventions for the patient.
In this project, we investigate the utility of convolutional neural networks in preforming fracture detection and classification of upper and lower extremity Xray images. We also explore the significance of dataset and preprocessing methods for our model performance.
This project is in collaboration of Zebra Medical Vision, for developing an algorithm to detect and localize fractures in musculoskeletal on radiograph images by deep learning techniques and reproduce state of the art results on collected large lower and upper extremity Zebra dataset.
Our model was implemented by a 50-layer residual neural network. We integrated some pre-processing algorithms, based on our knowledge for unique x-ray database. Finally, our model achieved appropriate performance results for our new Zebra dataset on shin and metacarpus classes, which are fitting to the known state of the art results.
From the results we suggested some dominant direction to improve model results, based on preprocessing features and characterization of unique Zebra dataset.
Our findings show that convolutional neural networks usage can be effectively for utilized for fracture detection and have a great potential for improving patient triage and diagnosis.

radiograph image

Recognition of Bone Fractures in X-Ray Images

Collaboration:
zebra logo
zebra logo