In this work we apply interpretable machine learning systems for the analysis of 18-FDG PET
brain scans from different medical centers. We show that the center origin of healthy control brain
images acquired with different scanners/protocols can be identified with high confidence. Consequently,
machine learning models trained with data from different scanners may be heavily impacted by this
bias when applied to a clinically relevant problem, e.g. the differential diagnosis of neurodegenerative
disorders. We propose IRMA (Iterated Relevance Matrix Analysis) as a recursive method to learn and
disregard bias in PET feature vectors. Related code is freely available at