Multi-dimensional neural network brain tissue quantification using frequency-sensitive magnetic resonance signal responses

• Applicant: Dr. Rahel Heule
• Subject Area: Medical Physics, Biomedical Technology
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 507678818

Quantitative MRI has proved to be closer to the origin of pathological tissue changes in neurological diseases than qualitative image contrasts, with higher sensitivity to detect abnormalities at early stages of the disease onset. To capture tissue characteristics related to microstructure, biochemical composition, or physiology as close as possible to the true complexity and differentiate between concurrent pathophysiological processes, multi-parametric methods are essential. Despite the recognized superiority of quantitative over qualitative weighted MR neuroimaging in terms of diagnostic information, multi-dimensional approaches, which are able to characterize multiple tissue-specific parameters simultaneously, have yet barely entered clinical routine. Multi-parametric quantitative imaging demands the acquisition of several differently weighted qualitative image contrasts, frequently leading to prohibitively long scan times, low spatial resolution, or limited anatomical coverage. We aim at exploiting the strong link to tissue microarchitecture and composition of frequency-sensitive MR signal responses generated by phase-cycled balanced steady-state free precession (bSSFP) imaging for multi-dimensional brain tissue characterization. To address the aforementioned limitations of state-of-the-art techniques, we propose to combine novel frequency-sensitive multi-contrast bSSFP sequence schemes with artificial neural network (ANN)-aided data analysis. In preliminary work, we have demonstrated that tissue parameters (T1, T2, diffusivity, fractional anisotropy) can be learned from the bSSFP frequency profile by supervised training. In this project, we will 1) maximize the quantitative information content extractable from multi-contrast bSSFP acquisitions targeting relaxation times (T1, T2, T2*), magnetization transfer (bound pool fraction, exchange rate), and/or diffusion metrics (diffusivity, fractional anisotropy, directionality information) and 2) improve the accuracy as well as increase the robustness of the parameter estimation to experimental noise, e.g. in case of highly accelerated acquisitions or at lower field strengths. Ultimate goal is to derive a rapid multi-dimensional 3D scan protocol at isotropic resolution with whole-brain coverage. The design of ANN architectures will include both supervised learning as well as unsupervised physics-informed training strategies driven by suitable signal models. From an MR acquisition perspective, emphasis will be put on effectively shaping the frequency response of bSSFP, e.g. by generating oscillating steady states, to yield high sensitivity to the targeted parameters. MR data for ANN training and testing will be collected in different signal-to-noise-ratio scenarios at low (64 mT), high (3 T), and ultra-high (9.4 T) field strengths.

• DFG Programme: Research Grants