Unsupervised anomaly detection in 2D radiographs using generative models

We present a method based on a generative model for detection of anomalies such as prosthesis, implants, screws, zippers, and metals in Two-dimensional (2D) radiographs. The generative model is trained following an unsupervised fashion using clinical radiographs as well as simulated data, neither of them containing anomalies. Our approach employs a reconstruction loss and a latent space consistency loss which have the benefit of identifying similarities which are forced to reconstruct X-rays without anomalies. In order to detect images with anomalies, an anomaly score is also computed employing the reconstruction loss and the latent space consistency loss. Additionally, the Frechet distance is introduced as part of the reconstruction loss. These losses are computed between an input X-ray and the one reconstructed by the proposed generative model. Validation was performed using clinical pelvis radiographs. We achieved an Area Under the Curve (AUC) of 0.77 and 0.83 with clinical and synthetic data, respectively. The results demonstrated a good accuracy of the proposed method for detecting outliers as well as the advantage of utilizing synthetic data for the training stage.