Deep networks such as the U-Net are outstanding at segmenting biomedical images when enough training data is available, but only then. Here we introduce a Domain Adaptation approach that relies on two coupled U-Nets that either regularize or share corresponding weights between the two streams, along with a differentiable loss function that approximates the Jaccard index, to leverage training data from one domain in which it is plentiful, to adapt the network weights in another where it is scarce.We showcase our approach for the purpose of segmenting mitochondria and synapses from electron microscopy image stacks of mouse brain, when we have enough training data for one brain region but only very little for another. In such cases, we outperform state-of-the-art Domain Adaptation methods.

Abstract. While Machine Learning algorithms are key to automating organelle segmentation in large EM stacks, they require annotated data, which is hard to come by in sufficient quantities. Furthermore, images acquired from one part of the brain are not always representative of another due to the variability in the acquisition and staining processes. Therefore, a classifier trained on the first may perform poorly on the second and additional annotations may be required. To remove this cumbersome requirement, we introduce an Unsupervised Domain Adaptation approach that can leverage annotated data from one brain area to train a classifier that applies to another for which no labeled data is available. To this end, we establish noisy visual correspondences between the two areas and develop a Multiple Instance Learning approach to exploiting them. We demonstrate the benefits of our approach over several baselines for the purpose of synapse and mitochondria segmentation in EM stacks of different parts of mouse brains.

No abstract