Leveraging Inter-Annotator Disagreement for Semi-Supervised Segmentation

The low signal-to-noise ratio typically found in biomedical images often leads experts to disagree about the underlying ground truth segmentation. While existing approaches for multiple annotations try to resolve conflicting annotations, we instead focus on efficiently using pixels of disagreement to estimate areas of high uncertainty in the data and exploit this information for semi-supervised segmentation. Pseudo-labelling approaches, which utilise unlabelled data by trying to match their own predictions, need to distinguish reliable from unreliable predictions. We propose to identify unreliable pseudo-labels from the output of a separate network that is trained to predict the uncertainty in the data based on conflicting annotations from different annotators. Compared to other uncertainty estimation techniques like MC-Dropout or ensembling approaches, our approach has the two key advantages that its estimate stems directly from the data and that it is computationally more efficient. Using two public datasets, we show the effectiveness of our approach.