Unsupervised pathology detection in medical images using learning-based methods

Detecting pathologies automatically is challenging because of their big variability. As the usual supervised machine learning approaches would only be able to detect one type of pathologies, in this work we pursue an unsupervised approach: learn the entire variability of healthy data and detect pathologies by their differences to the learned norm. Two methods have been developed based on this principle: A modified PatchMatch algorithm shows plausible results on contrasting brain tumors, but bad generalization ability for other types of data. A CVAE-based method on the other hand performs significantly better and ca. 17 times faster on the brain data and can be generalized to other pathologies, e.g. lung tumors. Not only is the achieved Dice coefficient of 0.55 comparable to other supervised methods on this data, moreover this method reliably detects different pathology types and needs no groundtruth.