Multi-Parametric Artificial Neural Network Fitting of Phase-Cycled Balanced SSFP Data

Representative multi-parametric reference and ANN output data for a healthy volunteer who was not included into neural network training (subject 3). ANN prediction of T1, T2, B1, and ΔB0 is applied to four different N-point bSSFP phase-cycling measurements: a standard unaccelerated protocol with N = 12 / R = 1 and three accelerated protocols with N = 12 / R = 2, N = 6 / R = 2, N = 4 / R = 2 (R: GRAPPA acceleration factor). The ANN outputs are compared to the reference parameter maps shown in the leftmost column. — Representative multi-parametric reference and ANN output data for a healthy volunteer who was not included into neural network training (subject 3). ANN prediction of T₁, T₂, B₁, and ΔB₀ is applied to four different N-point bSSFP phase-cycling measurements: a standard unaccelerated protocol with N = 12 / R = 1 and three accelerated protocols with N = 12 / R = 2, N = 6 / R = 2, N = 4 / R = 2 (R: GRAPPA acceleration factor). The ANN outputs are compared to the reference parameter maps shown in the leftmost column.

Representative multi-parametric reference and ANN output data for a healthy volunteer who was not included into neural network training (subject 3). ANN prediction of T₁, T₂, B₁, and ΔB₀ is applied to four different N-point bSSFP phase-cycling measurements: a standard unaccelerated protocol with N = 12 / R = 1 and three accelerated protocols with N = 12 / R = 2, N = 6 / R = 2, N = 4 / R = 2 (R: GRAPPA acceleration factor). The ANN outputs are compared to the reference parameter maps shown in the leftmost column.

Multi-parametric mapping techniques are an important tool in quantitative brain MRI to enable characterization of tissues and early diagnosis of pathologies within short acquisition times. Longitudinal and transverse relaxation times have proved to be a useful measure for monitoring brain tumors or neurodegenerative diseases, such as multiple sclerosis, Parkinson’s disease, or Alzheimer’s disease. Joint approaches, which allow quantifying multiple tissue parameters simultaneously based on the same acquisition scheme rather than performing sequential experiments for each, are particularly beneficial since they provide intrinsically co-registered maps.
In recent years, the concept of magnetic resonance fingerprinting (MRF) has become popular for simultaneous multi-parametric tissue characterization by generating signal evolutions, or fingerprints, thought to be unique for specific tissue types. Other novel approaches for combined T₁ and T₂ quantification utilize the mixed sensitivity to the longitudinal and transverse relaxation times of rapid non-balanced or balanced steady-state free precession (SSFP) sequences. Multi-pathway non-balanced SSFP imaging allows acquiring several steady-state configurations including higher order modes in a single scan, suited for simultaneous multi-parametric mapping. Phase-cycled balanced SSFP (bSSFP) acquisitions enable to sample the characteristic tissue-specific frequency profile, from which estimates for T₁ and T₂ can be derived.
The current phase-cycled bSSFP relaxometry techniques such as motion-insensitive rapid configuration relaxometry termed MIRACLE or the ellipse fitting approach termed PLANET are subject to a substantial mismatch between the estimated relaxation times of human brain tissues and measured gold standard reference values. This bias is especially prominent in white matter and likely linked to an asymmetric intra-voxel frequency content due to fiber tract geometry, in combination with multi-component relaxation, e.g. due to the presence of myelin. It has been reported that an asymmetric frequency distribution causes asymmetries in the bSSFP frequency profile. MIRACLE and PLANET fail to account for these anisotropies in the tissue microenvironment, but assume an idealized model for the bSSFP profile, i.e., a symmetric frequency response.
To eliminate this bias, in this work, a feedforward artificial neural network (ANN) is trained with phase-cycled bSSFP data acquired at 3 T in healthy volunteers as input and a multi-parametric output, i.e., T₁, T₂, transmit field, and off-resonance, using gold standard relaxation time measurements and reference field maps as target.

Rahel Heule, Jonas Bause, Orso Pusterla, Klaus Scheffler:

Multi-parametric artificial neural network fitting of phase-cycled balanced steady-state free precession data.

Magn Reson Med. 2020 Dec;84(6):2981-2993.