Doctoral candidate Burhaneddin Yaman was recently awarded the interdisciplinary doctoral fellowship for the 2020-2021 academic year. His research focuses on MRI reconstruction from sub-sampled data to improve patient care while also reducing scan time.
Continuing on that trajectory of excellence, Burhan also recently received the best paper award at the 2020 ISBI for his research on MRI reconstruction. Sponsored by the IEEE Signal Processing Society, and the IEEE Engineering in Medicine and Biology Society, the ISBI is a major conference on biomedical imaging. It is a competitive platform, and success at the symposium will help Burhan’s methodology gain wider traction and circulation. He is working towards his doctoral degree under the guidance of Prof. Mehmet Akçakaya. Burhan is also a recipient of several other awards: multiple ISMRM travel awards, an ECE department fellowship, the IEM Walter Lang Barnes travel fellowship, and an ISBI travel award. Supporting his doctoral work, are Burhan’s broad interests in the areas of signal processing, machine learning, and magnetic resonance imaging (MRI).
Burhan’s research mainly focuses on MRI reconstruction from sub-sampled data. Data acquisition during MRI is a slow process, one that involves trade-offs between resolution, scan time, and noise or signal-to-noise ratio. One way to improve the process without increasing the scan time, is to acquire less data (sub-sampling), and reconstruct the images using redundancies that are built into the data acquisition process such as multiple sensors, or redundancies in image structures.
A key part of Burhan’s doctoral work has to do with accelerated multi-dimensional MRI datasets suffering from high noise amplification. A critical outcome of this work is high-quality high-resolution images that are useful in a clinical context. The results have been shared at various conferences, and published in IEEE Transactions on Computational Imaging.
Another part of his research involves the application of machine learning to MRI reconstruction. Deep learning has recently emerged as an alternative technique for accelerated MRI because of its superior reconstruction quality as compared to conventional techniques. Although deep learning can be used as a black-box-tool for MRI reconstruction, it is beneficial to incorporate the information we know about the physics of MRI to this process, as it can ensure consistency with acquired measurements. Most current deep learning methods require fully sampled data to train artificial neural networks (ANN). But acquiring fully sampled datasets is typically impractical in the face of challenges such as organ movement, (a beating heart), signal decay during some MR scans, as well as long scan times during which the subject has to remain still.
To counter the problem, Burhaneddin has developed a self-supervised learning framework that enables training of physics-based deep learning MRI reconstruction without requiring fully-sampled data. He does this by splitting available data into two complementary disjoint sets, where one is used to train ANNs and the other is used to check the quality of the reconstruction.
The results so far have been promising, already demonstrating fourfold gains in scan time in both musculoskeletal and brain MRI. The results will soon be shared at multiple conferences, and a journal article detailing the work is set to be published in Magnetic Resonance in Medicine, a leading journal for MRI in medical applications.