Retrieval Optimization in Magnetic Resonance Fingerprinting (Master Thesis, Finished)

Author

Manuel Hürbin

Description

Magnetic Resonance Imaging has become one of the most important imaging methods in diagnosis and is an integral part of modern medicine nowadays. However, since the acquired images are often restricted to a qualitative characterization of the investigated anatomical structures, they do not provide any metrics to point towards a disease stage and early diagnoses are thus often difficult to make. In contrast, quantitative approaches have great potential in early disease detection, since they reflect properties and changes at the cellular level.

A new and promising quantitative approach introduced in 2013 is Magnetic Resonance Fingerprinting. This method basically measures the behavior of each pixel over a certain timespan, resulting in an individual signal evolution, represented as a vector. This signal alone does not suffice to reconstruct an image, but needs to be matched against a huge database with many simulated signals. The best match then yields the required information for a proper reconstruction of the image. Considering databases with over 250.000.000 entries, the time to find the best match is critical. Additionally, such large amounts of data require clever concepts to store and access data.

Since current research mainly uses MATLAB for the execution of the matching pipeline and mainly operates in volatile Random Access Memory, it is expected that this approach will reach its limits for large data sets and therefore alternatives are needed. This Thesis introduces Dactyloquant, which treats the aforementioned matching pipeline as a nearest neighbor search problem typically seen in multimedia retrieval. Dactyloquant uses and extends Cottontail DB - a specialized database for multimedia retrieval queries - for storage and data access. This newly developed system does not only scale well since it uses on-disk storage, it also introduces modern index structures, such as Super-Bit Locality-Sensitive Hashing or the Vector-Approximation+ File, that speed up the look up significantly and thus provide a valid alternative in current Magnetic Resonance Fingerprinting research.

Start / End Dates

2019/09/02 - 2020/02/29

Supervisors

Research Topics