Uncertainty-aware self-supervised neural network for liverT₁mapping with relaxation constraint.

Area of Science:

• Quantitative Magnetic Resonance Imaging (MRI)
• Artificial Intelligence in Medical Imaging
• Hepatology

Background:

• T1ρ mapping is a valuable non-invasive MRI technique for assessing liver tissue properties.
• Current learning-based T1ρ mapping requires extensive high-quality data and lacks uncertainty estimation.
• Accurate T1ρ quantification is crucial for diagnosing and monitoring liver diseases.

Purpose of the Study:

• To develop a learning-based liver T1ρ mapping method using fewer images.
• To incorporate uncertainty estimation into T1ρ quantification for enhanced reliability.
• To improve the efficiency and accuracy of liver T1ρ mapping in clinical settings.

Main Methods:

• A self-supervised neural network was developed, utilizing relaxation constraints for T1ρ mapping.
• Bayesian uncertainty estimation (epistemic and aleatoric) was integrated into the T1ρ quantification network.
• The method was validated on T1ρ data from 52 patients with non-alcoholic fatty liver disease.

Main Results:

• The proposed method achieved superior T1ρ quantification performance compared to existing techniques using only two T1ρ-weighted images.
• Integrated uncertainty estimation acted as a regularizer, enhancing model performance and correlating with confidence levels of liver T1ρ values.
• Experiments demonstrated the method's effectiveness in accelerating liver T1ρ mapping.

Conclusions:

• The developed learning-based approach significantly accelerates liver T1ρ mapping by reducing image acquisition requirements.
• Simultaneous uncertainty estimation provides crucial confidence metrics for clinical T1ρ quantification.
• This method holds significant potential for improving non-invasive liver disease assessment.