Quantitative Evaluation of Least-Squares Fitting Biases for Improvement of Diffusional Kurtosis Inference by Synthetic Q-space Learning

Abstract

Diffusional kurtosis imaging (DKI) quantifies the non-Gaussian diffusion properties of water molecules. For the inference of the DKI parameters (diffusion coefficient: D and diffusional kurtosis: K), the least-squares fitting (LSF) has been the standard technique. It is known that certain bias is observed especially in K value inferred by LSF depending on the signal-to-noise ratio (SNR) of diffusion-weighted images (DWI). For robust K inference, we developed synthetic Q-space learning (synQSL) technique that uses synthesized training data. The robustness of K inference by synQSL further improves based on additional correction by using LSF results as a reference. Therefore, it is considered that the bias suppression of K inference by LSF leads to improving the correction effect for synQSL. In this study, to improve the accuracy of the correction of K by synQSL, we evaluated the K inference biases by LSF using synthetic data and real image data of various SNR. As a result, we quantitatively showed that the inference bias depends not only on the SNR of the DWI, but also on the true value of K. This study is expected to improve the robustness of K inference by synQSL.