A Bayesian approach to diffusional kurtosis imaging.

Area of Science:

• Medical Imaging
• Biophysics
• Quantitative MRI

Background:

• Diffusional kurtosis imaging (DKI) metrics show promise as disease biomarkers due to their sensitivity to pathological changes.
• However, kurtosis maps often suffer from noise, compromising their visual quality and quantitative accuracy for clinical diagnosis.

Purpose of the Study:

• To develop and validate a Bayesian method for denoising kurtosis maps derived from DKI.
• To assess the method's ability to improve diagnostic performance and maintain biomarker utility without artificial parameter adjustments.

Main Methods:

• A Bayesian approach was employed, utilizing Gaussian priors derived from least-square method (LSM) estimations.
• The method automatically balances denoising with the preservation of metric dispersion, avoiding manual regularization parameter tuning.
• Performance was evaluated by comparing map quality against conventional methods and through receiver-operating characteristic (ROC) analysis for glioma grading using simulated diffusion-weighted imaging (DWI) data.

Main Results:

• The proposed Bayesian method significantly improved the quality of noisy kurtosis maps.
• Noninferiority to conventional methods was confirmed via area under the curve (AUC) tests for all kurtosis metrics.
• The method demonstrated superiority for several kurtosis-related metrics in differentiating glioma grades.

Conclusions:

• The Bayesian method enhances kurtosis map quality and preserves biomarker performance without increasing acquisition time or requiring arbitrary parameter selection.
• This approach facilitates the use of higher-order DKI fitting functions by mitigating overfitting, leading to improved estimation accuracy.