Chest tomosynthesis image enhancement by bone suppression using convolutional neural networks with synthetic data

Chest tomosynthesis (CTS) as a newly developed imaging technique can provide pseudo-3D volume anatomical information of the thorax from limited-angle projections. However, one drawback is that the acquisition angle for CTS imaging is much less than 180°, so the reconstructed image may be contaminated by structures behind and in front of the in-focus plane. To reduce the anatomical noise, in this study, we developed a deep learning-based bone suppression model to remove these occluding bony structures in CTS projections and the pseudo-3D images can then be reconstructed using the bone-suppressed projections. Specifically, the modified U-net architecture integrating with residual skip connections was utilized to perform bone suppression in CTS projections with multiple loss functions. The synthetic training data were generated by projecting the lung CT images with slice thickness of no more than 0.3 mm using the Hitachi CTS imaging geometry. To evaluate the bone suppression ability of the network trained with synthetic data in clinical practice, real patient data collected on the Shimadzu system were included. Experiments on both the simulation data and real patient data show that the proposed approach is effective in bone suppression while preserving detailed structures in CTS projections. The bone suppression rate (BSR) exceeded 90% for all cases. In the two simulation cases, the SSIM values were 0.9550 and 0.9484, with MAEs of 2.1930 and 2.3205, respectively. The reconstructed tomosynthesis images show that structures obscured by bone and difficult to identify were easier to observe after bone suppression. The bone-suppressed CTS images can be used as a supplement for radiologists in detecting lung lesions.