A comparison of fitting algorithms for diffusion-weighted MRI data analysis using an intravoxel incoherent motion model

The objective of this study is to propose a modified VARiable PROjection (VARPRO) algorithm specifically tailored for fitting the intravoxel incoherent motion (IVIM) model to diffusion-weighted magnetic resonance imaging (DW-MRI) data from locally advanced rectal cancer (LARC). The proposed algorithm is compared with classical non-linear least squares (NLLS) analysis using the Levenberg-Marquardt (LM) algorithm and with two recently proposed algorithms for 'segmented' analysis. These latter two comprise two consecutive steps: first, a subset of parameters is estimated using a portion of data; second, the remaining parameters are estimated using the whole data and the previous estimates. The comparison between the algorithms was based on the goodness-of-fit measure: performance analysis was carried out on real data obtained by DW-MRI on 40 LARC patients. The performance of the proposed algorithm was higher than that of LM in 64 % of cases; 'segmented' methods were poorer than our algorithm in 100 % of cases. The proposed modified VARPRO algorithm can lead to better fit of the IVIM model to LARC DW-MRI data compared to other techniques.