Purpose: Breast CT improves the visualization of breast anatomy over conventional 2D mammography and has been widely investigated for the diagnosis and treatment of breast cancer. Slot-scanning mammography has potential advantages with the use of high-resolution linear detectors and reduced x-ray scatter; however, integrating such technology into a CT acquisition introduces acquisition time challenges. In this work, we explore an implementation of slot-scanning CT which uses sparsely sampled projection data. We propose a dictionary-learning (DL) iteration reconstruction method to handle the sparse data and to facilitate fast breast CT scans using the slot-scanning approach.
Methods: We implemented a dictionary-learning penalized-weighted least squares (DL-PWLS) algorithm for breast CT. A global redundant dictionary with 16x16 voxel atoms was generated to emulate breast anatomical details. In each iteration, the image was extracted into patches that were estimated by a sparse combination of dictionary atoms. The DL-PWLS algorithm penalizes the difference between the reconstructed image and its representation assembled using the dictionary. The developed algorithm was tested using complete 360-view (49 kV, 243 mAs) projection data and 10% of data with 1D θ-sparsity (36 views) and 2D R,θ-sparsity (72 view but only 50% of the radial bins with 180° phase shifts between neighboring views).
Results: The DL-PWLS algorithm demonstrates enhanced noise reduction and low-contrast tissue classification over traditional filtered backprojection (FBP) and PWLS algorithms. For sparsely sampled data, visualizations of microcalcifications and soft tissues as measured by structural similarity (SSIM) are further improved using the DL-PWLS reconstruction with 10% 2D R,θ-sparsity (SSIM = 0.77 and 0.84 for microcalcifications and low-contrast tissue).
Conclusion: A DL iterative reconstruction algorithm is developed for breast CT. The DL-PWLS reconstruction in combination with doubly sparse data is expected to substantially shorten the breast CT scan time. The effective noise reduction also shows potential capability for patient dose reduction.
Funding Support, Disclosures, and Conflict of Interest: This work was supported, in part, by NIH grants R43CA224851 and R43CA239777, and an academic-industry partnership between Johns Hopkins University and Fischer Imaging.