MR Image Reconstruction and Inference – About Inverse Problems and Crimes
Deep learning tools have become an important ingredient of image reconstruction and inference. The increasing online availability of reference and annotated datasets has opened the field providing ample opportunity for a large community of researchers to develop and test new approaches. Yet, the underlying physics of how and which information is encoded may not readily be available, potentially leading to wrong assumptions about the data and in turn to data overfitting and overly optimistic results. To this end, the talk aims to revisit encoding and decoding of information in Magnetic Resonance (MR) imaging, the implications of bandlimitedness and data rasterization, the origin of noise and its statistics, data assumptions and data priors to encourage embedding of physics and signal theory into model learning and inference. Examples from cardiovascular MR will be used to illustrate concepts and an outlook into optimal experimental design for imaging tasks will be provided.