Clinical Measurements with Calibrated Instance-Dependent Confidence Interval

Abstract

Reporting meaningful confidence intervals for the predictions of a regression neural network is critical in medical imaging applications since clinical decisions are based on network predictions. We expect to obtain larger intervals for difficult examples and smaller ones for easier examples to predict. A recently proposed calibration procedure suggests predicting the mean and the variance and scaling the variance on a validation set. Another calibration approach is based on applying conformal prediction to quantile regression. We show that assuming a Gaussian distribution to predict the variance followed by a non-parametric Conformal Prediction technique to scale the estimated variance is the most effective way of achieving a small confidence interval with a coverage guarantee. We report extensive experimental results on various medical imaging datasets and network architectures.