Utilizing Conditions for Unsupervised Anomaly Detection in Brain MRI

Abstract

A need to label vast amounts of data in medical image analysis makes supervised algorithms time-consuming and raises concerns about incorrectly annotated pathologies. Unsupervised anomaly detection algorithms, which employ Generative Adversarial Networks (GANs), exist to tackle this issue. Such methods are supposed to detect unseen abnormal data by learning the distribution of the normal one. However, the norm is considered unified for a given task and does not account for any variability between samples, which may make its bounds vaguer. We assume that reckoning for external information about an image under examination can resolve this issue. This paper studies whether conditional GANs are suitable for patch-wise anomaly detection on brain MR images. We propose incorporating such attributes as age and patch position to better account for inter-patient variability. We train two GANs using \(64 \times 64\) images of chairs rotated by different angles from the RC-49 dataset and \(32 \times 32 \times 32\) patches from T1 weighted brain scans from the IXI dataset. We then reconstruct normal and abnormal samples with a modified image projection technique and use the obtained style vectors and the external attributes to assign anomaly scores to the images. On the test chair images, our approach achieves accuracy values of 88.4%, and we found it applicable to the 2D case. Nevertheless, on the brain patches, it shows a lower accuracy value of 64.3% for the test samples, indicating its inefficiency when applied to the 3D MR data in the proposed form. We also discuss the potential causes of the failed experiment and possible future avenues for improvement of the proposed approach.