Project DetailsIn this project we introduce a self-supervised, reference-free method for MRI pulse-sequence optimization, intended for use on both simulated and real MRI systems. Our core contribution is a contrastpromoting loss function that does not require any target contrast. Instead, our loss quantifies information-theoretic separability of signal intensities between tissues. Rather than relying on a target contrast, our method requires only an estimate of the scanned subject’s tissue structure, which our formulation allows to contain uncertainty and moderate errors. Lastly, our method allows to specify for each optimization instance which tissues are desired to be more distinct than others, by a weighting scheme. Optimizations with our method uncover a low-dimensional manifold of optimized contrasts, on which a specific requested tissue emphasis, thus weights, selects a unique contrast.
