Conditional Least Squares Estimation of Diffusion MRI Parameters

Jelle Veraart¹, Sofie Van Cauter², Stefan Sunaert², Jan Sijbers¹

Many diffusion models require highly diffusion-weighted images, which suffer from low signal-to-noise ratio, eddy current distortions, and subject motion. Prior to diffusion model fitting, those distortions should be corrected; That data correction alters the underlying Rician data distribution. Therefore, previously proposed methods remove the Rician bias might be suboptimal. We propose a conditional least squares estimator (CLS), which is from theoretically an equivalent to the maximum likelihood estimator (MLE). However, the CLS remains, unlike the MLE, in well-defined cases unbiased after data correction as it can rely on the linearity of the expectation operator in high SNR and homogeneous regions.