Feature density as an uncertainty estimator method in the binary classification mammography images task for a supervised deep learning model

Abstract

Labeled medical datasets may include a limited number of observations for each class, while unlabeled datasets may include observations from patients with pathologies other than those observed in the labeled dataset. This negatively influences the performance of the prediction algorithms. Including out-of-distribution data in the unlabeled dataset can lead to varying degrees of performance degradation, or even improvement, by using a distance to measure how out-of-distribution a piece of data is. This work aims to propose an approach that allows estimating the predictive uncertainty of supervised algorithms, improving the behaviour when atypical samples are presented to the distribution of the dataset. In particular, we have used this approach to mammograms X-ray images applied to binary classification tasks. The proposal makes use of Feature Density, which consists of estimating the density of features from the calculation of a histogram. The obtained results report slight differences when different neural network architectures and uncertainty estimators are used