10:00 AM - 11:00 AM
Learning to Reconstruct without Ground Truth Data
RSL Future Leaders Seminar Series
Lengthy data acquisition times remain a main challenge in MRI. Recently, deep learning (DL) techniques have emerged as an alternative accelerated MRI technique due to its improved reconstruction quality compared to conventional parallel imaging and compressed sensing approaches. Among DL approaches, physics-guided DL (PG-DL) approaches have drawn interest as it incorporates the domain knowledge by unrolling an iterative algorithm for solving a regularized reconstruction. PG-DL approaches are typically trained in a supervised manner using fully-sampled data as ground-truth reference. However, it is either challenging or impossible to acquire fully-sampled data in a number of cases such as dynamic/high-resolution MRI, hindering applicability of PG-DL methods.
In this talk, we present a novel self-supervised learning technique for PG-DL MRI reconstruction in the absence of fully-sampled data. The proposed method splits the acquired measurements into two disjoint subsets, one of which is utilized to enforce data consistency within the unrolled network, while the other is used to define loss in k-space. We show that the proposed self-supervised technique trained on sub-sampled data performs similar to the supervised approach trained on fully-sampled data. We also discuss extensions of the proposed approach for further improvements and other application.
RSL Group Meeting ZOOM Link
Burhaneddin Yaman is a Ph.D. student in Electrical engineering and Computer Science at the University of Minnesota, working with Prof. Mehmet Akcakaya. He has received a B.Sc. in Electronics and Communication Engineering from Istanbul Technical University, Turkey, in 2016. His research interest includes machine learning, magnetic resonance imaging and tensor decompositions.