Calibration of A Regression Network Based on the Predictive Variance with Applications to Medical Images

Calibrating regression neural networks is crucial in medical imaging applications where the decision making depends on the predicted confidence. Modern neural networks are not well-calibrated and they tend to overestimate confidence when compared to the expected accuracy. In this study we propose a calibration procedure for regression networks that is based on scaling the predictive variance. The amount of scaling depends on the input and is a function of the predictive variance computed by the network. We thus find an entire calibration function instead of a single parameter. We report extensive experiments on a variety of image datasets and network architectures. Our approach achieves state-of-the-art results with a guarantee that the prediction accuracy is not altered.