Amplifying image quality gain in x-ray phase contrast imaging of mastectomy samples with deep learning denoising.
Authors
Affiliations (13)
Affiliations (13)
- School of Physics, The University of Melbourne, School of Physics, Melbourne, Victoria, 3010, AUSTRALIA.
- Melbourne Data Analytics Platform, The University of Melbourne, Melbourne Data Analytics Platform, Melbourne, 3010, AUSTRALIA.
- Faculty of Health Sciences, The University of Sydney, Faculty of Health Sciences, Sydney, 2006, AUSTRALIA.
- Maroondah BreastScreen, Eastern Health, Ringwood East, Victoria, 3134, AUSTRALIA.
- Monash Health, Monash Health, Clayton, 3168, AUSTRALIA.
- Australian Synchrotron, ANSTO, 800 Blackburn Road, Clayton, Victoria, 3168, AUSTRALIA.
- AS Science, ANSTO, 800 Blackburn Road, Clayton, Victoria, 3168, AUSTRALIA.
- ANSTO, AS Science, Kirrawee, 2232, AUSTRALIA.
- Australian Synchrotron, ANSTO, AS Science, Sydney, 2232, AUSTRALIA.
- Manufacturing, CSIRO, Private Bag 10, Clayton South, Victoria 3169, Clayton South, Victoria, 3169, AUSTRALIA.
- School of Physics, The University of Melbourne, School of Physics, Melbourne, 3010, AUSTRALIA.
- University of Sydney Faculty of Health Sciences, University of Sydney Faculty of Health Sciences, Lidcombe, New South Wales, 2141, AUSTRALIA.
- Department of Chemistry, University of Melbourne, Melbourne, Victoria 3010, Melbourne, 3010, AUSTRALIA.
Abstract
Phase-contrast computed tomography (PCT) of the breast has previously been shown to produce higher-quality images at lower radiation doses without the need for breast compression. The present study is aimed at further reduction of the radiation dose in PCT, while preserving or further increasing the image quality, by applying supervised deep learning denoising of reconstructed PCT images. This work was carried out in preparation for live patient PCT breast cancer imaging, initially at specialised synchrotron facilities.
Approach: PCT scans of 34 fresh full mastectomy samples were acquired using propagation-based phase-contrast imaging with 32 keV monochromatic parallel X-rays at mean glandular doses of 4 mGy and 24 mGy. All scans were reconstructed using Filtered Back Projection algorithm with Paganin's phase retrieval. A supervised U-Net-based deep learning denoising model was trained on 28 pairs of 4 mGy and 24 mGy scans and then applied to denoise the remaining 6 stacks of reconstructed 4 mGy images. Denoised PCT images were quantitatively evaluated using signal-to-noise ratio (SNR), spatial resolution, structural similarity index measure (SSIM) and peak-signal-to-noise ratio (PSNR). The images were also visually compared and systematically assessed by experienced medical imaging specialists and radiologists. 
Main Results: Deep learning denoising increased SNR by a factor of four while spatial resolution remained unchanged. SSIM and PSNR improved from 0.89 and 37 dB to 0.96 and 42 dB, respectively. Visual assessors significantly preferred the denoised images over the original 4 mGy images, and visual assessment indicated no increase in perceived artefacts in denoised images compared with the original 4 mGy images.
Significance: Deep learning-based image denoising can further improve image quality in PCT without increasing radiation dose in imaging of mastectomies, supporting the feasibility of lower-dose PCT protocols or improved image quality for future clinical applications.