Comparative Analysis of Diffusion Tensor Imaging Estimation Methods

Somaye   Jabari; Amin  Ghodousian; Reza  Lashgari; Babak A.  Ardekani

doi:10.18502/fbt.v13i2.21943

Somaye Jabari Department of Algorithms and Computation, Faculty of Engineering Science, College of Engineering, University of Tehran, Tehran, Iran
Amin Ghodousian Department of Algorithms and Computation, Faculty of Engineering Science, College of Engineering, University of Tehran, Tehran, Iran
Reza Lashgari Institute of Medical Science and Technology, Shahid Beheshti University, Tehran, Iran
Babak A. Ardekani Center for Advanced Brain Imaging and Neuromodulation, The Nathan S. Kline Institute for Psychiatric Research, Orangeburg, New York, USA

DOI: https://doi.org/10.18502/fbt.v13i2.21943

Keywords: Diffusion Tensor Imaging; Diffusion Magnetic Resonance Imaging; Tensor Estimation Method; Cholesky Decomposition.

Abstract

Purpose: Diffusion Tensor Imaging (DTI) is a noise-sensitive method, where a low Signal-to-Noise Ratio (SNR) results in significant errors in the estimated tensor field. This topic focuses on a comprehensive evaluation of various DTI estimation methods, such as Linear Least Squares (LLS), Weighted Linear Least Squares (WLLS), iterative re-weighted Linear Least Squares (IRLLS), and Non-linear Least Squares (NLS). The article will explore how each method performs in terms of accuracy, efficiency in estimating the diffusion tensor and robustness against noise.

Materials and Methods: The study compares the methods using simulated diffusion-weighted Magnetic Resonance Imaging (MRI) data. Time complexity and performance of the LLS, WLLS, IRLLS, and NLS methods were evaluated across key metrics such as TRMSE, RMSE, MSD, and ΔSNR.

Results: The results of the study demonstrate that LLS and IRLLS consistently outperform other methods in terms of TRMSE, MSD, and SNR, particularly in high-noise scenarios. NLS performs best in reducing RMSE but high noise causes it to fit to noise, so it is not robust. WLLS showed the weakest performance across all metrics.

Conclusion: The paper suggests that LLS, despite its simplicity, remains a competitive option in terms of capturing the true underlying diffusion properties. IRLLS further refines this by iteratively reducing the effect of outliers in tensor estimation.