Uncertainty Measurements for the Reliable Classification of Mammograms

Abstract

We propose an efficient approach to estimate the uncertainty of deep-neural network classifiers based on the tradeoff of two measurements. The first based on subjective logic and the evidence of soft-max predictions and the second, based the Mahalanobis distance between new and training samples in the embedding space. These measurements require neither modifying, nor retraining, nor multiple testing of the models. We evaluate our methods on different classification tasks including breast cancer risk, breast density, and patch-wise tissue type and considering both an in-house database of 1600 mammographies, as well as on the public INBreast dataset. Throughout the experiments, we show the ability of our method to reject the most evident outliers, and to offer AUC gains of up to 10%, when keeping 60% of most certain samples.