PROJECT TITLE :

Visual Correspondences for Unsupervised Domain Adaptation on Electron Microscopy Images

ABSTRACT:

For Electron Microscopy volumes, we provide an Unsupervised Domain Adaptation approach. Pretrained models are able to work on new data because to our way of aggregating visual correspondences (motifs that are visually similar across acquisitions). Annotations from an existing acquisition are examined to identify pivot points that characterise the reference segmentation, and patch matching algorithms are used to detect their visual counterparts in a new volume. This heatmap shows how frequently places on the new volume match relevant areas from the initial acquisition, and it is generated by a voting mechanism that collects all of the candidate correspondences. As a result of this information, we can perform model adaptations in two different ways: either by a) optimising model parameters under a Multiple Instance Learning formulation, so that predictions between reference locations and their sets of correspondences agree, or by b) using high-scoring regions of the heatmap as soft labels to be incorporated in other domain adaptation pipelines, including Deep Learning ones. Both mitochondria and synapses can be segmented using unsupervised techniques that are qualitatively consistent with results achieved with complete supervision, and this can be done without putting in any additional annotation effort.