Domain Adaptation of MRI Segmentation
Visualizing representation-space cracks to evaluate segmentation quality
As a research assistant at TU Delft (Imaging Physics), I worked on deep-learning domain adaptation for MRI segmentation, MRI image processing, and manifold learning.
Based on the observation of organ-related cracks in the visualization of a segmentation network’s representation spaces (hidden-layer outputs), I proposed a visualization method of these cracks to evaluate segmentation quality, and showed that UNet-like models tend to generate them via the skip-connection mechanism.