[Evaluation of Diffusional Kurtosis Inference Using Synthetic q-space Learning and Bias Correction]

Area of Science:

• Diffusion MRI
• Machine Learning
• Biomedical Imaging

Background:

• Synthetic q-space learning (synQSL) is a deep learning approach for inferring diffusional kurtosis (K).
• A known bias in synQSL is dependent on the noise level of synthetic training data.
• Bias correction is crucial for accurate K estimation in diffusion MRI.

Purpose of the Study:

• To evaluate the accuracy of K inference using synQSL with bias correction.
• To compare synQSL with bias correction against traditional least-squares fitting (LSF) methods.
• To assess the robustness and error reduction achieved by bias-corrected synQSL.

Main Methods:

• K was inferred using synQSL on synthetic and real diffusion MRI data.
• Bias correction was applied to synQSL-inferred K values.
• Results were compared to K inferred by LSF.
• Inference robustness was assessed using outlier rates and root mean square error (RMSE) across varying noise levels and number of excitation (NEX).

Main Results:

• SynQSL without correction demonstrated a lower outlier rate than LSF across noise levels.
• Bias correction further reduced the outlier rate in synQSL.
• The smallest RMSE was achieved with bias-corrected synQSL, particularly when comparing NEX 1 to NEX 4 data.

Conclusions:

• SynQSL combined with bias correction offers a robust method for K inference.
• This approach significantly reduces errors compared to LSF.
• Bias-corrected synQSL is a promising technique for accurate diffusional kurtosis estimation in diffusion MRI.