You are here: Home NEWS & EVENTS Latest News
Latest News
  • [ October 27, 2020]

    Qiushi Ren’s group from Peking University makes new progress on multi-material decomposition in spectral computed tomography

  • Recently, Prof. Qiushi Ren’s group (Department of Biomedical Engineering, College of Engineering, Peking University) made new progress on multi-material decomposition in spectral computed tomography (CT). A novel deep-learning based method was proposed and carefully examined, which has been published on the journal, IEEE Transactions on Medical Imaging, entitled “PMS-GAN: Parallel Multi-Stream Generative Adversarial Network for Multi-Material Decomposition in Spectral Computed Tomography”.

    CT, as a non-invasive anatomic imaging modality, plays an essential role in clinical diagnosis. Nowadays, techniques such as photon-counting detectors (PCD) and fast kVp-switching have enabled spectral CT, which is capable of measuring attenuation at different energy regimes and differentiating materials. Material decomposition, as an important application of spectral CT, aims to extract the constituent material distribution from the CT data acquired at different X-ray energy levels. However, most analytical material decomposition methods are susceptible to the inevitable nonlinear behavior and ill-posed of CT systems, which results in low decomposition accuracy. Furthermore, when the number of decomposed materials or energy bins increases, the decomposition task will become more complicated.

    For the purpose of decomposing multiple basis materials as well as increasing the decomposition accuracy, Ren’s group proposed a novel deep learning network named parallel multi-stream generated adversarial network (PMS-GAN). The parallel multi-stream structure reduced the interference among sub-generators; and the novel differential map (DM) was designed to better estimate the concentrations of various basis materials at individual pixel positions. In addition, a more explicit loss function was designed based on Markovian discriminator, which boosts the performance of the network.

    Figure 1. The structures of PMS-GAN. (a) the entire schema of PMS-GAN. (b) the schema of generators when the number of basis materials is 3. (c) the detailed architecture of the first generator in (b).

    PMS-GAN was examined in both simulation and phantom experimental studies compared with the state-of-the-art methods. In the simulation study, contrast agent (Ultravist370), bone, and bone marrow were chosen as the objective materials, while in the experimental study, PMS-GAN was investigated on decomposing biopsy needles and torso phantom. Visually, PMS-GAN suppressed noise and the residual of non-objective materials. Quantitatively, PMS-GAN outperformed the other reference methods in structural similarity index and Pearson correlation coefficient. The study demonstrates promising generalization capability of PMS-GAN and proves that it has certain potential toward clinical applications.

    Figure 2. The decomposition results of PMS-GAN and reference methods in simulation data.

    This research was supported by the Natural Science Foundation of China, Shenzhen Science and Technology Program, and the Natural Science Foundation of Hebei Province.