Projection Domain Metal Segmentation With Iterative Epipolar Consistency

The main task of my bachelor’s thesis was binary segmentation of high-density areas in the 2D projection domain of CT scans, which was needed for some downstream task.

Additionally, I experimented with approaches that aimed at increasing segmentation performance by increasing epipolar consistency between different angles’ segmentations. Recurrently building up 3D context within the 2D U-Net turned out to be too resource-intensive to get to batch sizes that would allow for stable training. Instead, I introduced a novel postprocessing step that was able to increase epipolar consistency and thereby segmentation performance somewhat, also at the cost of increased computational needs.

The project was completed as a cooperation with Siemens Healthineers, which owns the IP to my bachelor’s thesis. That’s why I’m keeping this project page so vague, and do not include a download link for the full thesis.