Zlatko Sokolikj - Dissertation Defense
"Redistilling GRAPPA: A Bipartite Graph Formulation for Parallel MRI Reconstruction"
Abstract:
Magnetic resonance imaging (MRI) is a powerful diagnostic modality, widely valued for its ability to provide detailed, noninvasive insights into the human body, while acquisition speed remains fundamentally linked to the sampling of k-space. Parallel MRI addresses this limitation by leveraging multiple receiver coils with spatially varying sensitivity profiles to reconstruct undersampled data. Among these approaches, GRAPPA has emerged as a widely adopted autocalibrated method, learning local k-space interpolation weights from a fully sampled autocalibration signal (ACS) region. While highly effective, its reliance on a fixed local kernel structure can limit adaptability, particularly under higher acceleration, where region-specific variations become more pronounced.
This dissertation introduces a graph-based perspective on GRAPPA-style reconstruction, motivated by the observation that the interpolation kernel admits a natural interpretation as a weighted bipartite graph. Within this framework, acquired multi-coil k-space samples define source nodes, missing samples correspond to target nodes, and learned interpolation coefficients encode weighted edges between them. The resulting formulation, termed Bipartite-Graph GRAPPA (BG-GRAPPA), preserves the scan-specific, ACS-based calibration of GRAPPA while making source–target connectivity explicit and more amenable to interpretation.
