Robust DTI Noise Level Estimation Improves RESTORE Tensor Estimation

Introduction: Noise level estimation in DTI plays a crucial role in determining outliers, fitting likelihood-based tensor models, and estimating the reliability of measured quantities. However, noise level estimation methods developed for other magnetic resonance techniques are often inappropriate for DTI acquisitions. For example, noise level cannot be directly computed from background regions with protocols using background suppression (e.g., CLEAR) or exhibiting spatially varying noise due to parallel imaging or variable coil sensitivity (e.g. SENSE). DTI images are commonly up-sampled by zero-padding the Fourier coefficient images, so the local noise structure is highly correlated. Thus, noise level estimation from local regions is difficult. Spatial variability and correlation limit the applicability of both the single image (based on background intensities) and double image (based on moments of Rician random variables) methods [1]. When the noise level specified for the RESTORE tensor estimation method is too low, too much data is excluded and the error rate suffers [2]. Conversely, when the noise level is specified too high, artifacts are not excluded and the error rate increases. Thus, accurate noise level specification is crucial. We develop a new estimation approach based on noise invariance to diffusion weighting and demonstrate that this technique improves the RESTORE method of outlier rejection and tensor estimation.