Applying Multimedia Retrieval Approaches to Speed-up Magnetic Resonance Fingerprinting (Master Thesis, Ongoing)
Magnetic Resonance Fingerprinting (MRF) is a novel approach that aims at speeding up the process of quantitative medical imaging without impacting the quality of the result, with the overall objective of facilitating early diagnoses. For this, MRF measures a unique signal (fingerprint) for each pixel of the image. This fingerprint is then matched against a database of simulated signals to recover the parameters which the observed signal originated from, and several quantified maps are reconstructed.
In a previous project, it could be shown that approaches from multimedia similarity search are very effective for the matching step of the MR fingerprinting pipeline. For this, the multimedia database Cottontail DB has been extended by Locality-Sensitive Hashing (LSH), an approximate index structure. While LSH has significantly reduced retrieval time, the approximation has also led to a deterioration of the quality of the retrieval results and therefore of the entire MR fingerprinting approach.
The objective of the preparation phase of this Master’s thesis is to explore, in collaboration with the Department of Biomedical ENgineering, how the first very encouraging results in the application of multimedia retrieval techniques for MR fingerprinting can be improved and extended. This includes, but is not limited to:
• Extension of the Cottontail database by exact index structures (e.g. VA, VA+, etc.)
• Extension of the Cottontail database by additional approximate index structures (e.g., Product Quantization, etc.)
• Adapting existing as well as new index structures to the data by (for example) identifying a structure in the dictionary that can be exploited
• Definition of progressive query processing for the MR fingerprinting matching process. Progressive querying jointly considers several index structures; first results will be produced by approximate index structures and allow for presenting first results to a physician; these first results will be gradually refined when more exact index structures return
• Addressing scalability of the entire retrieval approach to allow for larger MRF databases
• Identification of criteria for the automated optimization of the matching/retrieval process by selecting the best suited index structures / query execution
• Evaluating the overall approach using different database sizes and query models
Start / End Dates
2020/05/08 - 2020/12/07