HDNLS: Hybrid Deep-Learning and Non-Linear Least Squares-Based Method for Fast Multi-Component T1ρ Mapping in the...

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Quantitative Imaging
• Biophysics

Background:

• Non-linear least squares (NLS) is standard for T1ρ mapping but is slow and sensitive to initial guesses.
• Deep learning (DL) methods are faster but can be noise-sensitive and require NLS data for training.
• Existing methods struggle to balance speed, accuracy, and noise robustness in quantitative MRI parameter estimation.

Purpose of the Study:

• To develop a hybrid deep learning and non-linear least squares (HDNLS) model for accelerated multi-component T1ρ parameter mapping.
• To evaluate the performance of HDNLS and its variants for T1ρ mapping, particularly in the knee joint.
• To investigate the impact of NLS iterations as a regularization technique within the HDNLS framework.

Main Methods:

• Developed HDNLS, combining voxel-wise DL trained on synthetic data with iterative NLS.
• Introduced four HDNLS variants (Ultrafast-NLS, Superfast-HDNLS, HDNLS, Relaxed-HDNLS) to balance speed and accuracy.
• Analyzed the effect of NLS iterations on HDNLS performance and parameter estimation stability.

Main Results:

• HDNLS achieved comparable accuracy to NLS and regularized-NLS (RNLS) with a significant speed improvement (minimum 13-fold).
• HDNLS demonstrated superior estimation quality compared to pure DL methods while maintaining high speed.
• The number of NLS iterations in HDNLS effectively regularizes parameter estimation, improving robustness to noise.

Conclusions:

• HDNLS provides a fast, reliable, and accurate solution for multi-component T1ρ mapping, overcoming limitations of NLS and DL alone.
• The HDNLS framework offers tunable configurations to meet specific requirements for speed and performance in quantitative MRI.
• HDNLS represents a significant advancement for efficient and robust parameter estimation in T1ρ imaging applications.