Nature Sciences

New Progress in the Reconstruction of Neural Tissue Structure Made by BIT

  Recently, Ye Chuyang, associate professor of the School of Information and Electronics, Beijing Institute of Technology(BIT), and Li Yuxing, a doctoral student, collaborated with Deputy Chief Physician Zeng Xiangzhu, Department of Radiology, Peking University Third Hospital used a deep learning-based method to improve the reconstruction quality of neural tissue structure under the condition of limited number of diffusion gradients, and realized the uncertainty quantification of the reconstruction results based on diffusion magnetic resonance imaging and the method based on deep learning. Relevant results were published in the title of "An improved deep network for tissue microstructure estimation with uncertainty quantification" in Medical Image Analysis, a top journal in the field of medical image processing (impact shadow IF 8.88).

Figure 1 Reconstructure error of neural tissue structure

  Nerve tissue structure information measured by diffusion magnetic resonance imaging has been widely used in neuroscience research. This information can prompt brain development and aging and is also related to many neurological diseases, providing important biomarkers for neuroscience research. However, in a typical imaging scenario, due to the limitation of imaging time, the accuracy of neural tissue structure reconstruction was affected. In addition, the uncertainty information of the reconstruction result was also important for subsequent image analysis, but the existing methods cannot provide relevant information.

Figure 2 Schematic diagram of neural tissue structure reconstruction and uncertainty quantification results

  In order to solve these problems, the research team proposed an improved deep network. While adaptively combining historical information, the spatial domain and angle domain dictionary were used to sparsely represent the signal, and then the signal sparse representation was mapped to the neural tissue structure reconstruction result in a separable form. Besides, this method was based on the sparse representation form and used the Lasso Bootstrap strategy to quantify the uncertainty of the neural tissue structure reconstruction results. The team used the public brain diffusion magnetic resonance imaging data set to verify the designed method on different signal models. Its accuracy is significantly better than the existing methods, and it has obtained meaningful quantitative results of uncertainty. This work provides important new ideas for the calculation of neuroscience-based on diffusion magnetic resonance imaging.



Release Date: 2020-05-06

Contribution: School of Information and Electronics

Editor: Cao Anqi

Translator: Leng Junbo, News Agency of BIT