Low-Dose 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) Department of Biomedical Imaging, University of Arizona, Tucson, AZ
Presentations
PO-GePV-I-18 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: Deep learning has emerged as a prominent approach for low-dose computed tomography (CT). Most deep learning methods require a large amount of labeled data for supervised training. This can be challenging or infeasible for medical imaging applications. In this work, we investigated an unsupervised learning-based denoising from a single image called Self2Self for low-dose breast CT.
Methods: A sample case (300 views, full-scan, 10.9 mGy) from a HIPAA-compliant breast CT dataset of women assigned BIRADS 4/5 was used retrospectively with IRB approval. The projection dataset was uniformly undersampled to provide 150 views (5.5 mGy) to emulate sparse-view acquisition. Both full-view and sparse-view projection datasets were reconstructed using the Feldkamp-Davis-Kress (FDK) algorithm, and a compressed sensing-based algorithm, FRIST. The full-view or sparse-view FDK reconstruction was fed as the only training sample to a Self2Self network. Each network was trained for 1E5 iterations. The pixel-wise dropout probability for Self2Self was varied from 0.2 to 0.7 and the resulting images were visually analyzed to determine the best choice. PSNR was computed with the full-view FRIST reconstruction as the reference.
Results: For full-view (300 views), Self2Self reduced the noise in the FDK images and improved PSNR by 8 dB. The microcalcification was well-retained in the Self2Self image. For sparse-view (150 views), the FDK, Self2Self, and FRIST reconstructions yielded PSNR of 30.9 dB, 39.3 dB, and 36.5 dB, respectively. Sparse-view FDK suffered from streaking artifacts smearing along the breast periphery. Using the sparse-view FDK image as the input, Self2Self showed minor residual streaks in the reconstruction and was comparable to sparse-view FRIST reconstruction.
Conclusion: The Self2Self denoising framework for sparse-view breast CT could enable dose reduction. This single-image based Self2Self learning approach does not require labeled, diverse, training data. Future work will investigate incorporating a physical model for reducing streak artifacts in sparse-view reconstructions.
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