Short-Scan Cone-Beam Dedicated Breast CT: Self-Supervised Denoising From Single Image
Z Fu1*, H Tseng1, A Karellas1, S Vedantham1,2, (1) Department of Medical Imaging, University of Arizona, Tucson, AZ, (2) Biomedical Engineering, University of Arizona, Tucson, AZ
Presentations
PO-GePV-I-2 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: Supervised learning methods have emerged for medical imaging reconstruction or image denoising. However, it can be impractical to obtain a large, diversified, training dataset of input and target pairs. Further, supervised neural networks generalize poorly on out-of-distribution data. In this work, we explore an unsupervised learning approach called Self2Self to denoise breast CT images for short-scan acquisitions.
Methods: HIPAA-compliant breast CT datasets of BIRADS 4/5 women acquired on a prone, clinical-prototype, dedicated cone-beam breast CT system were used retrospectively with IRB approval. The full-scan (300 views/360 degrees) projection data (mean glandular dose, MGD: 11.3 mGy) was retrospectively sampled to provide 225 views (270 degrees) short-scan data (MGD: 8.5 mGy). The full-scan and the short-scan data (with Parker weights) were reconstructed (0.273 mm voxels) using the ramp-filtered Feldkamp-Davis-Kress (FDK) algorithm. The full-scan data were also reconstructed using a fast ASD-POCS-based algorithm named FRIST. The FDK reconstructed short-scan slice was inputted as the only training sample to a randomly initialized Self2Self network. The Self2Self network was trained for 1E5 iterations using a workstation with 8 GB GPU and took approximately 2.5 hours for a 512x512 image. PSNR was computed using the full-scan FRIST reconstruction as reference.
Results: Compared to input FDK image, Self2Self reduced image noise, yielding 1.5 dB improvement in PSNR. Nevertheless, the artifacts present in the FDK short-scan reconstruction remained in the Self2Self reconstruction but were suppressed in the FRIST reconstruction. Self2Self yields 2.2 dB lower PSNR compared to the FRIST short-scan reconstruction.
Conclusion: We validated the Self2Self denoising framework for short-scan breast CT. Unlike data-based supervised learning methods, this single-image-based Self2Self learning approach has no prerequisite on training data. Further work is needed to incorporate physical models for reducing artifacts in addition to image denoising and will be pursued in the future.
Funding Support, Disclosures, and Conflict of Interest: Supported in part by National Cancer Institute (NCI) of the National Institutes of Health (NIH) grants R01 CA241709, R01CA199044 and R21 CA134128. The contents are solely the responsibility of the authors and do not necessarily reflect the official views of the NCI or the NIH.
Keywords
Taxonomy
IM- Cone Beam CT: Machine learning, computer vision
Contact Email