Physical imaging parameter variation drives domain shift

Statistical learning algorithms strongly rely on an oversimplified assumption for optimal performance, that is, source (training) and target (testing) data are independent and identically distributed. Variation in human tissue, physician labeling and physical imaging parameters (PIPs) in the generative process, yield medical image datasets with statistics that render this central assumption false. When deploying models, new examples are often out of distribution with respect to training data, thus, training robust dependable and predictive models is still a challenge in medical imaging with significant accuracy drops common for deployed models. This statistical variation between training and testing data is referred to as domain shift (DS).To the best of our knowledge we provide the first empirical evidence that variation in PIPs between test and train medical image datasets is a significant driver of DS and model generalization error is correlated with this variance. We show significant covariate shift occurs due to a selection bias in sampling from a small area of PIP space for both inter and intra-hospital regimes. In order to show this, we control for population shift, prevalence shift, data selection biases and annotation biases to investigate the sole effect of the physical generation process on model generalization for a proxy task of age group estimation on a combined 44 k image mammogram dataset collected from five hospitals.We hypothesize that training data should be sampled evenly from PIP space to produce the most robust models and hope this study provides motivation to retain medical image generation metadata that is almost always discarded or redacted in open source datasets. This metadata measured with standard international units can provide a universal regularizing anchor between distributions generated across the world for all current and future imaging modalities.