Deep learning for MRI-based CT synthesis: a comparison of MRI sequences and neural network architectures

Synthetic computed tomography (CT) images derived from magnetic resonance images (MRI) are of interest for radiotherapy planning and positron emission tomography (PET) attenuation correction. In recent years, deep learning implementations have demonstrated improvement over atlas-based and segmentation-based methods. Nevertheless, several open questions remain to be addressed, such as which is the best of MRI sequences and neural network architectures. In this work, we compared the performance of different combinations of two common MRI sequences (T1- and T2-weighted), and three state-of-the-art neural networks designed for medical image processing (Vnet, HighRes3dNet and ScaleNet). The experiments were conducted on brain datasets from a public database. Our results suggest that T1 images perform better than T2, but the results further improve when combining both sequences. The lowest mean average error over the entire head (MAE = 101.76 +/- 10.4 HU) was achieved combining T1 and T2 scans with HighRes3dNet. All tested deep learning models achieved significantly lower MAE (p < 0.01) than a well-known atlas-based method.