Purpose: To develop a deep learning (DL) model for scatter correction of dedicated breast CT (BCT) projection images.
Methods: 118 digital breast phantoms were obtained by automatic segmentation of as many BCT patient images and augmented via random translation of the breast in the image field of view to a total of 470 samples. The samples were used to simulate the primary and scatter projections of a clinical BCT system using Monte Carlo (MC) simulations. Twelve projections were simulated for each sample, one every 30° covering a full 360° acquisition. The projections simulated from 450 samples were used to train the DL model to predict the scatter from the input primary+scatter projections. The model was trained end-to-end by unifying information from four orthogonal projections, to account for breast size and position. It was then tested on the remaining samples, and on acquired projections of a physical breast phantom. The scatter-corrected projections of this phantom were reconstructed, and then compared to the uncorrected reconstruction and a reconstruction from MC-corrected projections.
Results: Inside the breast, the mean percentage error (min, max) between the MC-simulated and the DL-predicted scatter was -1% (-10%, 4%) and the mean standard deviation of the error within a projection was 6% (2%, 13%). In a homogeneous region of the physical phantom reconstruction, the mean ΔHU between ROIs close to the perimeter and the center were 86, 24, -64 for the uncorrected, MC-corrected, and DL-corrected images, respectively.
Conclusion: The proposed DL model showed promising results for scatter correction in BCT images. Moreover, its performance allows for almost real-time processing, correcting 300 projections in 17 seconds on a standard PC. Future work will involve the evaluation of different DL architectures and learning algorithms to further improve the precision of the method, and additional evaluation with physical phantoms and patient images.