Raw-Data-Based Material Decomposition Using Modified U-Net for Low-Dose Spectral CT

Basis material decomposition has been a common tool in spectral computed tomography (CT), which is used to acquire the material-specific information about human tissues. However, the performance of current methods for material decomposition is seriously affected by noise, especially when the X-ray radiation dose is reduced. To address this problem, a novel approach with modified U-net convolutional neural network is proposed and evaluated. First, the raw data (sinograms) collected from spectral CT detector are decomposed into different substances directly. Then a filtered back projection algorithm is used to generate the basis material maps. The modified U-net is trained with the raw data in the projection domain, and the input of the network are different energy sinograms while the decomposed sinograms of basis materials could be obtained from the output. In the experiment, water and iodine are selected as basis materials. Comparing with the traditional method, the PSNR of materials maps is improved about 7dB to 17dB of different slices, which shows that this proposed approach could make more accurate material maps in low-dose spectral CT.