Comparing a deterministic and a Bayesian classification neural network for chest diseases in radiological images

Abstract

A common mantra for automated decision systems is that a system should know when it doesn’t know. Bayesian neural networks are designed to capture uncertainties over the network weights and in theory, they perform better predictions and output uncertainties. To this end, we compare in this paper a deterministic neural network and a Bayesian neural network for the classification of chest diseases in radiological images. We use the ChestX-ray14 data set involving 14 respiratory diseases like pneumonia and atelectasis. We found that the deterministic network similar to CheXNet outperformed the Bayesian version in this task, whereas, employed on the more simplistic MNIST dataset it did not. Our experiments suggest that there is a gap between theory and practical use of BNNs for very deep networks and real clinical data.