This master thesis was completed in the X-Ray Tomography Group and focused on improving the quality of breast cancer CT scans. The project involved using phase micro CT scanning to obtain high-resolution images of breast samples, with the goal of being able to identify tiny tumor features in breast tissues. While CT stands for computed tomography, which is a technique that uses computer algorithms to reconstruct 3D images from projection data, the scanning of phase is the measurement of ray deviation rather than amplitude. Phase measurments provides contrast for tissue identification. However, phase CT scanning can take days to complete, which is too long for practical use in breast cancer diagnosis.
To address this issue, we used deep learning to achieve higher quality scans in a shorter amount of time. The project involved developing a scanning and processing pipeline to handle volumetric data and using the Noise2Noise procedure to train an artificial intelligence with lower quality inputs and regress the higher quality signal contained in the pair of inputs. Alignment techniques such as software registration and modifications to the scanning procedure were used to achieve the best results with the input pairs. The results showed that deep learning quality enhancement outperformed conventional filtering techniques and could potentially improve breast cancer diagnosis in conjunction with classical histological diagnosis.