Myo-regressor Deep Informed Neural NetwOrk (Myo-DINO) for fast MR parameters mapping in neuromuscular disorders

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biophysics

Background:

• Muscle Magnetic Resonance Imaging (mMRI) parameter mapping traditionally uses computationally intensive pattern recognition algorithms.
• Existing deep learning (DL) models for MR parameter mapping lack interpretability and haven't been applied to mMRI for Neuromuscular Disorders (NMDs).

Purpose of the Study:

• To develop an efficient and explainable DL model, Myo-DINO (Myo-Regressor Deep Informed Neural NetwOrk), for mapping Fat Fraction (FF), water-T2 (wT2), and B1 parameters in NMDs.
• To integrate physics-based constraints into a DL framework to enhance model interpretability and performance.

Main Methods:

• A Physics-Informed Neural Network (PINN) with a U-Net architecture (Myo-DINO) was developed using 2165 MESE slices from 232 NMD subjects.
• Two physics-informed loss functions were implemented: Cycling Loss 1 (mono-exponential model) and Cycling Loss 2 (Extended Phase Graph (EPG) theory with slice profile).
• Myo-DINO was trained with varying weights for the L2 norm and physics-informed components, including a self-supervised approach.

Main Results:

• Myo-DINO achieved superior performance with Cycling Loss 2, demonstrating high reconstruction similarity (SSIM > 0.92, PSNR > 30.0 dB) and low error (NRMSE < 0.038) compared to reference maps.
• Muscle-wise FF, wT2, and B1 predictions showed good agreement with reference values.
• The self-supervised constraint with Cycling Loss 2 improved explainability by ensuring the network learned according to EPG model assumptions.

Conclusions:

• Myo-DINO provides a robust and efficient workflow for mMRI parameter mapping in NMDs, serving as a viable alternative to existing post-processing algorithms.
• The Extended Phase Graph (EPG) model incorporated in Cycling Loss 2 offers the most effective physical constraints for this multi-parameter regression task.
• Physics-informed DL, particularly with EPG constraints and self-supervision, enhances both the efficiency and interpretability of MR parameter mapping in NMD research.