Deep Learning Framework for Sparse Tomographic Image Reconstruction

Researchers at Stanford have developed a methodology for deep learning-based image reconstruction by incorporating the physics or geometry priors of the imaging system with deep neural networks. This invention may be implemented in different imaging modalities such as MRI and CT. The researchers are the first to investigate the problem of volumetric MRI using ultra-sparse k-space samples that can be acquired within a second. Their physics-aware deep learning framework is able to reconstruct high quality MRI from sparse k-space samples, and they have validated the proposed approach across various abdominal patients. For CT applications, the researchers have investigated the novel-view projection synthesis problem for X-ray imaging. Their approach can also be generalized to a more general synthesis from multi-views to multi-view projections. Their solution provides a deep learning-based geometry-integrated projection synthesis model to generate novel-view X-ray projections through feature disentanglement and geometry transformation. They have validated their approach using X-ray projections across various lung patients.