Recently, the research team led by Zhang Guanglei from the Advanced Innovation Center for Biomedical Engineering published their research results entitled “3D Deep Encoder–Decoder Network for Fluorescence Molecular Tomography” in Optics Letters, a world-renowned journal in optics and photonics community. The team demonstrates a novel Fluorescence molecular tomography (FMT) reconstruction method with greatly improved image quality and significantly reduced reconstruction time based on the deep learning technology. In their research, an end-to-end three-dimensional deep encoder–decoder (3D-Encoder–Decoder) network is proposed to improve the quality of FMT reconstruction, which marks the first successful attempt to apply Artificial Intelligence (AI) to FMT reconstruction. The research is supported by the National Natural Science Foundation of China, the Programme of Introducing Talents of Discipline to Universities and the Advanced Innovation Center for Biomedical Engineering.
Model architecture: the 3D deep encoder–decoder (3D-En–Decoder) network
Recent years witness a rapid development of smart health care with the threshold of an “AI plus” era. Some key technologies like deep learning boost a new round of development of AI, which further promotes the integration of AI and the medical industry with data-intensive, knowledge-intensive and mental-work-intensive features. Two enormous challenges researchers face in biological and medical science are the precision cancer detection at early stage and the long-term observation of cancer cells at molecular and cellular levels. The noninvasive, in vivo FMT can track cell and molecular activities including the occurrence, development and transfer of cancer cells in real time. However, the poor image quality and long reconstruction time are main hindrances to the application of FMT to Fluorescence molecular tomography (FMT) biological and medical researches.
Unlike traditional methods in which the FMT reconstruction problem is explicitly defined and domain-knowledge is carefully engineered into the solution, this network does not benefit from prior knowledge like optical parameters, but instead directly establishes the nonlinear mapping relationship between the inside fluorescent source distribution and the boundary fluorescent signal distribution. The parameters of nonlinear mapping are studied and adjusted continuously in the process of network training. With this technique, any inaccuracy caused by establishing the photon propagation model or solving the ill-posed inverse problem can be fundamentally avoided, since the forward and inverse problems do not need to be explicitly solved anymore. Better image contrast and localization accuracy were obtained by using the proposed network. In particular, the new reconstruction speed increases by more than 1,000 times compared to that of the old method.
The paper is available at:
Reported by Lin Ouya and Yang Xiaoyu
Edited by Jia Aiping
Reviewed by Liu Yan
Translated by Xiong Ting