Colloquium with Zlatko Sokolikj (2026-09-30)

"Redistilling GRAPPA: A Bipartite Graph Formulation for Parallel MRI Reconstruction"

Wednesday, Sep 30, 2026

Colloquium - 499 DSL Seminar Room: 03:30 to 04:30 PM Eastern Time (US and Canada)

Abstract:

Magnetic resonance imaging (MRI) produces high-resolution images of soft tissue without ionizing radiation, but acquisition remains relatively slow because the scanner must sample raw spatial-frequency data, known as k-space, before reconstructing the final image. Parallel MRI accelerates this process by using multiple receiver coils to recover information from undersampled data. GRAPPA is a widely used k-space parallel imaging method that estimates missing samples through interpolation weights learned from a fully sampled autocalibration signal (ACS) region.

This talk presents Bipartite-Graph GRAPPA (BG-GRAPPA), a graph-based view of GRAPPA reconstruction in which acquired multicoil k-space samples are represented as source nodes, missing samples are represented as target nodes, and calibrated interpolation weights form the edges between them. This reformulation preserves the scan-specific calibration structure of classical GRAPPA while making its source-to-target connectivity explicit.

Using this graph view, the talk examines how reconstruction behavior changes across anatomy, acceleration, coil relationships, target classes, and spatial location in k-space. The results show that reconstruction difficulty is not uniform: performance varies across k-space, and improvements in magnitude image quality do not always correspond to improved complex k-space agreement. By making GRAPPA’s interpolation structure visible, BG-GRAPPA provides a framework for analyzing which connections matter, where calibration transfers reliably, and how more adaptive reconstruction strategies may be designed.

Research Image