(Sunday, 7/14/2019) 4:00 PM - 5:00 PM
Room: 303
Purpose: To denoise projection data obtained on a grating-based phase-contrast micro-computed tomography (μCT) system using a convolutional neural network (CNN), and quantify image quality improvement in the attenuation, phase, and dark-field tomographic reconstructions.
Methods: A plastic container with reservoirs of various liquids was scanned on a phase-contrast Talbot-Lau grating-based μCT system. Object and reference projection data sets of size 256 x 256 pixels with 6 phase steps were obtained for 360 angles. A CNN with 6 convolutional layers was trained to reduce both Poisson and salt-and-pepper noise. Training data were generated by adding Poisson-distributed noise and randomly masking 2% of the input pixels to mimic the salt-and-pepper noise found in the original images. These noise maps were randomly generated throughout the training procedure, ensuring that no two training images were exactly alike. 20 random projection angles were reserved to monitor for overfitting. The projection data was run through phase retrieval. Attenuation, phase, and dark-field images were reconstructed using filtered back projection. This process was carried out for original and CNN-denoised data sets both with and without median filtering. The resulting images were compared using RMS measurements, line profiles, and visual inspection.
Results: Compared to the original data, the CNN-denoised data showed decreased RMS noise by factors of 1.87, 1.28, and 3.3 in the attenuation, phase, and dark-field reconstructions, respectively. Similar noise reductions were observed for the median filtered comparison. Line profiles of the original and CNN-denoised data show no spatial resolution loss. The CNN-denoised data improves delineation of low-contrast structures, particularly in the dark-field.
Conclusion: This technique demonstrates the effectiveness of using a CNN to correct for both Poisson and salt-and-pepper noise in the projection domain for phase-contrast μCT. The corrected projections result in reduced noise in the attenuation, phase, and dark-field reconstructions while maintaining edge sharpness and increasing conspicuity.
Funding Support, Disclosures, and Conflict of Interest: Funding provided by a Mayo/ASU Team Science Award and the Mayo Clinic Radiology Research Scholarship Fund. Dr. McCollough receives grant funding from Siemens Healthcare for work unrelated to this project.
Keywords
Not Applicable / None Entered.
Taxonomy
IM- CT: Machine learning, computer vision
Contact Email
