A single convolutional neural network for simultaneous estimation of breast thickness map and scatter maps in dual-energy digital breast tomosynthesis using a dual-layer detector

Recently, a direct-indirect dual-layer flat-panel-detector (DLFPD) has been proposed and applied to dual-energy digital breast tomosynthesis (DEDBT) for spectral imaging. The DLFPD acquires spatio-temporally aligned low-energy (LE) and high-energy (HE) images with a single exposure. However, both LE and HE images suffer from severe scattered radiation due to the dominant scatter contribution from HE x-rays, degrading image quality and lowering lesion contrast. Scattered radiation is highly dependent on breast thickness, which is also crucial for optimizing image post-processing and analysis, such as weighted subtraction for DE images and volumetric breast density calculation. In this work, we propose a single convolutional neural network to simultaneously estimate LE and HE scatter maps, as well as the breast thickness map, by leveraging their relationship and shared information. The network was trained and evaluated with Monte Carlo simulated projection images of anthropomorphic digital breast phantoms with varying glandularity and compressed thickness, which cover the range typically seen in clinical exams. The global mean absolute relative error for LE and HE scatter estimation was below 4% across all projection images of testing phantoms. The global mean absolute error for breast thickness map estimation was approximately 1.1 mm, with high accuracy in the central region and a faithful prediction of thickness roll-off in the peripheral region. After scatter correction, cupping artifacts were noticeably reduced, and lesion contrast and detectability were significantly improved.