Magnetic Resonance Fingerprinting Reconstruction using Methods from Multimedia Retrieval (Master Thesis, Finished)
Magnetic resonance fingerprinting (MRF) is a relatively new form of magnetic resonance imaging (MRI) with the capability of simultaneous quantitative mapping of multiple properties. A relatively short, pseudorandom data acquisition is followed by a computationally expensive reconstruction step in which aqcuired signals are compared to a large dictionary of simulated signals to retrieve tissue parameters. The computational cost associated with fingerprint matching imposes constraints on dictionary size and ultimately on the number of measurable parameters and their resolution.
In this project, we introduce concepts and methods from approximate nearest neighbor search to reduce the computational cost for MRF matching. Our main contribution consists of extending product quantization (PQ) based maximum inner product search (MIPS) to the complex domain. We achieved 91% correctly matched fingerprints (recall) at 12 % the time required for the exhaustive search during the evaluation of our preliminary implementation with in vivo data at 1.5 T. Using phantom samples at low field (100mT), we verified that the incurred error is low compared to the overall accuracy of the MRF method and could potentially outperform existing approaches to improving MRF reconstruction optimization.
This project was realized in a collaboration with the Adaptable MRI Technology Center (AMT-Center) at the Department of Biomedical Engineering of the University of Basel.
Start / End Dates
2020/05/08 - 2020/12/07