Abstract:
Unsupervised Domain Adaptation compensates for domain shifts on Electron Microscopy volumes. Our method aggregates visual correspondences—motifs that are visually similar across acquisitions—to infer changes on pretrained model parameters and enable them to operate on new data.
In particular, we use a patch matching algorithm to find pivot locations that characterize the reference segmentation in a new volume. We create a consensus heatmap by voting on all candidate correspondences and mapping how often locations on the new volume match relevant locations from the original acquisition.
This information allows us to adapt models in two ways: a) optimizing model parameters under a Multiple Instance Learning formulation to match predictions between reference locations and their sets of correspondences, or b) using high-scoring regions of the heatmap as soft labels for other domain adaptation pipelines, including deep learning ones.
We demonstrate that these unsupervised techniques can produce high-quality segmentations on unannotated volumes, qualitatively consistent with full supervision, for both mitochondria and synapses, without new annotation effort.
Note: Please discuss with our team before submitting this abstract to the college. This Abstract or Synopsis varies based on student project requirements.
Did you like this final year project?
To download this project Code with thesis report and project training... Click Here
