The consideration of predictive uncertainty in medical imaging with deep\nlearning is of utmost importance. We apply estimation of both aleatoric and\nepistemic uncertainty by variational Bayesian inference with Monte Carlo\ndropout to regression tasks and show that predictive uncertainty is\nsystematically underestimated. We apply $ \\sigma $ scaling with a single scalar\nvalue; a simple, yet effective calibration method for both types of\nuncertainty. The performance of our approach is evaluated on a variety of\ncommon medical regression data sets using different state-of-the-art\nconvolutional network architectures. In our experiments, $ \\sigma $ scaling is\nable to reliably recalibrate predictive uncertainty. It is easy to implement\nand maintains the accuracy. Well-calibrated uncertainty in regression allows\nrobust rejection of unreliable predictions or detection of out-of-distribution\nsamples. Our source code is available at\nhttps://github.com/mlaves/well-calibrated-regression-uncertainty\n