Unsupervised scalable anomaly detection: application to medical imaging