Anomaly detection and segmentation methods using Variational Autoencoders in Brain MRI

Abstract

Variational Autoencoders (VAEs) have emerged as a promising technique for unsupervised anomaly detection and segmentation in brain MRI. The principle behind this deep generative modeling is learning a model of healthy brain anatomy by encoding this information into a latent representation and reconstructing it. Anomalies can be detected and localized by discrepancies between the reconstructed data and the original input. This technique relieves the need for pixel-level segmented data and provides the possibility to detect arbitrary anomalies. Challenges have been reported in recent publications regarding the sensitivity towards bright lesions and the limitations of commonly used anomaly scoring methods based on reconstruction error and residual images. These issues are addressed in this work by changing the background of brain MR scans, and introducing anomaly scoring methods based on outlier detection, and activation maps. The evaluation of these methods is performed using four synthetic datasets containing toy anomalies varying in size, intensity, and number. The results show the possibility of detecting dark lesions, the potential of outlier detection in latent space for anomaly detection, and the extraction of activation maps for anomaly segmentation.