Metal artifact reduction with deep learning based spectral CT

Computed tomography (CT) has been intensively used for medical diagnostics because of its high temporal and spatial resolution and low cost. However, metal objects (e.g. implants) in presents cause metal artifacts which may severely deteriorate image quality and impede diagnosis. This work proposed a deep learning based spectral CT method for correcting metal artifacts caused by implants. The method first used a convolutional neural network (CNN) to synthesize a triple-energy spectral CT dataset from a dual-energy dataset, and then sent the dataset to a second CNN to synthesize the virtual monochromatic image. The proposed method was validated through simulations of an abdomen phantom. Results demonstrated that our method can effectively reduce metal implant induced streaks in the adjacent regions. The CT number root mean square error (RMSE) was under 30 HU.