Synthesizing human brain CT images from MRI with deep learning

Rapid development of artificial intelligence techniques significantly promotes important medical imaging applications such as disease diagnosis, medical image analysis and cross-modality synthesis. It is valuable to synthesize Computed Tomography (CT) scans of human brains from Magnetic Resonance Imaging (MRI) scans because CT images contain particular information for diseases such as malignant tumors, cerebrovascular malformations and aneurysms. Although there has been many studies trying to apply machine learning methods (e.g., fully convolutional neural networks and generative adversarial networks) to predict CT images based on MRI, there are still opportunities to increase the quality of synthesized CT images using the latest machine learning algorithms. In this study, we aim to synthesize high-credibility human brain CT images from MRI images using machine learning approaches. We build a dataset of both CT and MRI images of 41 patients from The Cancer Imaging Archive. A data pre-processing pipeline is designed to guarantee that the processed images are suitable as inputs to machine learning algorithms. The pipeline includes (1) pairing MRI and CT scans according to a certain time interval between CT and MRI exams of the same patient, (2) registering all MRI images to a standard MRI template, (3) registering all CT images to paired MRI images, (4) intensity normalization and (5) extracting 2D slices from 3D CT and MRI volumes. Three deep neural networks with encoder–decoder structure are selected and implemented including U-Net, U-Net++ and pix2pix to synthesize human brain CT images from MRI images. Experimental results illustrate that U-Net++ shows better performance than U-Net and pix2pix in synthesizing CT images from MRI by achieving a Mean Squared Error of 0.025, a Mean Absolute Error of 0.082 and a Peak Signal-to-Noise Ratio of 67.90, respectively. The trained machine learning models from this work can be applied in future human brain CT image prediction where CT images are absent and MRI images are available for specific patients, thereby supporting disease diagnosis based on CT images.