Jiulong Liu,National University of Singapore
Room 306, No.5 Science Building
We propose a learnable method for the l1-norm related regularization of inverse problems for medical image reconstruction whose solutions are known to exhibit artifacts and deteriorated image quality when the data is insufficient. Our method allows the learnable prior information embedded in our framework to extremely alleviate the ill-conditioness in the process of inversion and anisotropically approximate nature distribution compared to pre-defined prior. The method is applied to many examples, learned shape prior for low-dose CT reconstruction and fast MRI reconstruction, jointly learned absorption coefficients and scattering coefficients prior for DOT reconstruction, learned spatiotemporal prior for 4DCT reconstruction, etc.