Brain Tissue Mapping and Segmentation by MRI-based Blind-Source-Separation Techniques

Brain or any other tissue signatures are considered to be linear combinations of tissue components. Mixtures of such tissue components are "blindly" separated by means of geometrical sparse component analysis. The original set of at least two MR images, acquired by using the spin-echo or spoiled FLASH techniques with specific set of Tr and Te, are sparsified by using multiple wavelets and curvelets. The algorithms and techniques are investigated by separating simulated MRI images, where the ground truth is available. They are then applied to clinical data. Both iterative FCM and robust regression lend themselves to good estimation of the mixing matrix and thereby separation of the tissue components. Further improvements are discussed.