Show abstract - PDF - Reviews

The main benefit of unsupervised anomaly detection is the ability to identify arbitrary, rare data instances of pathologies even in the absence of training labels or sufficient examples of the rare class(es). In the clinical workflow, such methods have the potential to assist in screening and pre-filtering exams for potential pathologies and thus meaningfully support radiologists. Even though much work has been done on using auto-encoders (AE) for anomaly detection, there are still two critical challenges to overcome: First, learning compact and detailed representations of the healthy distribution is cumbersome. Recent work shows that AEs can reconstruct some types of anomalies even better than actual samples from the training distribution. Second, while the majority of unsupervised algorithms are tailored to detect hyperintense lesions on FLAIR brain MR scans, recent improvements in basic intensity thresholding techniques have outperformed them. Moreover, we found that even state-of-the-art (SOTA) AEs fail to detect several classes of non-hyperintense anomalies on T1w brain MRIs, such as brain atrophy, edema, or resections. In this work, we propose reversed AEs (RA) to generate pseudo-healthy reconstructions and localize various brain pathologies. We extensively evaluate our method on T1w brain scans and increase the detection of global pathology and artefacts from 73.1 to 89.4 AUROC and the amount of detected local pathologies from 52.6% to 86.0% compared to SOTA methods.
Hide abstract