My bachelor’s thesis’ main task was binary segmentation of high-density areas in the 2D domain of CT scans. The primary goal was to cut down the total time required to produce the final 3D CT reconstruction while keeping similar reconstruction quality. Advantages of reduced CT reconstruction times are manifold. For instance, it allows for a higher frame rate in CT-navigated surgery.
The project was completed as a cooperation with Siemens Healthineers, which owns the IP to my bachelor’s thesis. That’s why I have to keep this project page a bit vague, and cannot include a download link to the full thesis.
The core principle of CT reconstruction is that 2D images are recorded by the CT device at different angles and streamed into the reconstruction software sequentially. 2D processing steps are performed concurrently using a pipeline architecture, and only at the end a 3D image can be reconstructed using the processed 2D images.
The segmentation of high-density areas in 2D is required to reduce image artifacts that occur after performing the final reconstruction to 3D. Existing segmentation algorithms are only able to work in the 3D domain, which means that the full process effectively requires two reconstructions: Reconstructing from 2D to 3D, segmenting in 3D, projecting the mask back to 2D, using the 2D masks for some processing step, doing the final reconstruction back to 3D.
My work enabled performing the segmentation in 2D. Consequently, only a single reconstruction to 3D is required, and the total processing time is cut down significantly.
In addition to training the segmentation model itself, I introduced a novel 2D post-processing step that improves mask quality by increasing epipolar consistency between a mask and the masks of previously processed 2D images.
Of course, since the project was done as a bachelor’s thesis, I was not able to do everything from collecting a dataset to integrating my approach into the final product on my own. But I managed to deliver a proof of concept. To my knowledge, based on my results, the Siemens Healthineers team started collecting a more extensive dataset and integrating my approach into the product.