Autoencoders for unsupervised anomaly segmentation in brain MR images: A comparative study

Deep unsupervised representation learning has recently led to new approaches in the field of Unsupervised Anomaly Detection (UAD) brain MRI. The main principle behind these works is learn a model normal anatomy by compress and recover healthy data. This allows spot abnormal structures from erroneous recoveries compressed, potentially anomalous samples. concept great interest medical image analysis community as it i) relieves need vast amounts manually segmented training data—a necessity for pitfall current supervised Learning—and ii) theoretically detect arbitrary, even rare pathologies which might fail find. To date, experimental design most hinders valid comparison, because they are evaluated against different datasets pathologies, use resolutions iii) architectures with varying complexity. intent this work establish comparability among recent methods utilizing single architecture, resolution same dataset(s). Besides providing ranking methods, we also try answer questions like how many subjects needed normality if reviewed sensitive domain shift. Further, identify open challenges provide suggestions future efforts research directions.