A Weighted Consensus Deep Learning Method for the Fusion of Object Segmentation Models and its Application to Glomeruli of the Mouse Olfactory Bulb

Several deep learning-based object segmentation models have been proposed for biomedical image analysis, each with different strengths and weaknesses. Here, we propose a weighted consensus method for the fusion of object segmentations. For each object, the semi-supervised learning method follows the credibilities to reach a consensus on the determination of the correct segmentation from different algorithms. The user can train the weights of different algorithms in the consensus network to predict a new fused segmentation that combines the advantages of the individual object segmentation algorithms. We apply this consensus model to the segmentation of glomeruli in the mouse olfactory bulb with improved performance over a single-model approach. This weighted consensus deep learning method can also be applied to the fusion of other object segmentations. We develop a widget in Napari that supports a user-friendly GUI for the convenient application of this human-in-the-loop method.