Project Details
Projekt
Data science and pattern recognition
Subject Area
Medical Informatics and Medical Bioinformatics
Term
since 2023
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 500888779
Multi-contrast and hyperspectral imaging have been employed in various fields, including magnetic resonance imaging (MRI). In MRI, multi-contrast imaging has been used to classify tissues based on relaxation properties. Here, machine learning (ML) can support the differentiation of tissues with regards to granularity and robustness. MRI fingerprinting uses multi-echo signals to predict T1 or T2 times to characterize tissues, but the characterization of molecular composition (MR biosignatures) using this method is unknown. Techniques such as Diffusion Tensor Imaging and q-space trajectory imaging in MRI measure molecular motion in tissues and are commonly used to characterize sub-voxel-sized structures, CEST and X-nuclei MRI enable further complementary measurements. There is a need for approaches that can combine multiple MR contrasts and extract sub-voxel information in order to fully characterize tissues.The objective of Project D is to identify and understand tissue- and pathology-specific magnetic resonance (MR) biosignatures through the use of ML-based analysis of a large set of MR signatures acquired in the projects in area A and C. This will involve using ML techniques for pixel-accurate image fusion to combine multiple contrasts, multi-contrast tissue classification and segmentation, interactive labeling, and the derivation of acquisition parameters for optimized scan protocols. For image analysis, the project will explore relevant image contrasts and acquisition parameters using attention-based approaches. For interactive labeling, the project will develop techniques to derive low-dimensional representations of the hyperspectral (multi-contrast) data. This low-dimensional representation will allow for the visualization and visual analysis of clusters, and manual corrections will be made to update the high-dimensional distances. Lastly physics-informed deep learning will be used to explore to what extent the acquisition protocols can be improved regarding acquisition time using end-to-end optimization. Based on the knowledge derived in these work packages, the project will inform the optimization of scanning protocols for the second phase of the RU. The developed tools will be made available to the S project. The research will be conducted in two phases, with the first phase focusing on the development of tools for data analysis and the second phase using these tools to gain insights. The data for the research will be collected from both healthy subjects (A projects) and subjects with clinical questions (C projects) and will be coordinated through the S project. Close collaboration with the corresponding projects will ensure physically correct models and pathology-relevant experiments as well as early dissemination of the insights and developed tools.
DFG Programme
Research Units
