Improving Pathological Distribution Measurements with Bayesian Uncertainty.

Deep learning assisted histopathology has the potential to extract reproducible and accurate measurements from digitised slides in a scalable fashion. A typical workflow of such analysis may involve instance segmentation of relevant tissues followed by feature measurements. Inherent segmentation uncertainties produced by these deep models, however, could propagate to the downstream measurements, causing biased distribution estimate of the whole slide. One challenging aspect when handling ambiguous tissues is that the number of instances could differ as the instance segmentation step may not generalise well to these tissues. As an attempt to address this problem, we propose to derive a confidence score from the segmentation uncertainties obtained from Bayesian Neural Networks (BNNs) and utilise these as weights to improve the distribution estimate. We generate a synthetic dataset that mimics the diverse and varying visual features of the original data to enable systematic experiments. With this dataset we demonstrate the robustness of the method by extracting several clinically relevant measurements with two different BNNs. Our results indicate that the distribution estimates are consistently improved when the instances are weighted by the entropy-derived confidence measure. In addition, we provide results on applying the method to the original data.