Deep Learning Approaches for Contrast Removal from Contrast-enhanced CT

In internal radiation therapy, dosimetry is essential to predict its efficacy and potential side effects. Contrast enhanced computed tomography (ceCT) is most commonly used as starting point for planning. However, native CT (nCT) is required for accurate dosimetry computations. In thiswork,we propose an in-silico method to remove the contrast agent from ceCT images so that the Hounsfield Units (HU) would be similar to those in nCT. Two approaches, one paired-image neural network (NN) and one un-paired NN, were applied to ceCT/nCT image pairs for contrast removal.We evaluated their performance in terms of HU values, and performed dosimetry calculations on the original nCT and ceCT, and on the in-silico nCTs to evaluate the impact on the dose rate. The two approaches yielded good results both in terms of HU reduction (more than 30%) and in the difference of dose rate against the original nCT (less than 1.38% vs. 4.76%).