Intracellular volume fraction (ICVF) estimates from diffusion MRI (dMRI) based biophysical models can be a valuable biomarker for pathologies which cause neuronal damage or degradation [1]. As one of the main factors in the recently proposed in vivo g-ratio estimation [2], which could conceptually be linked to axonal conduction velocity and thus axonal function, the extracted intracellular volume fraction estimates must be accurate and reproducible. With recent biophysical diffusion models being devised for specific purposes, we examine the in vivo agreement of publicly available methods (CHARMED, DKI, and NODDI [3-6]) in single and multi-fibre populations throughout the brain.Test-retest multi-shell dMRI data was acquired on a 3 T GE MR750 with 11 b=0-images and shells at b=300, 700, and 2500 s/mm² with 8, 32, and 64 DWIs per shell, respectively. FOV of 24×24 cm was acquired with a 96×96 matrix and 50 slices of 2.5 mm thick. SENSE=2, TE/TR=71.7/5200 ms, and scan time of 9m58s. A 3D-T1-weighted IR-FSPGR was also acquired (1 mm isotropic resolution, TE/TR/TI: 3.1/7.4/400 ms). To determine regions of single and multiple fibre populations, constrained spherical deconvolution (CSD) was performed in ExploreDTI [7] with L_MAX=8 and to get the number of fibre populations (as in [8]). CHARMED was fit using default parameters which include two restricted (axonal) compartments (16 fitted parameters), initialising ICVF to 0.3 (as in [3]). Kurtosis tensors were fit in ExploreDTI and the axonal water fraction (AWF [5]) was extracted (21 fitted parameters). NODDI was fit on all data with default parameters, allowing a single dispersing fibre population (5 fitted parameters). ICVF extracted from NODDI was corrected for the isotropic volume fraction (similar to [9]). ICVF were examined separately for regions of one, two, and three fibre populations for DKI, NODDI, and CHARMED. Coefficients of variation (CV) were defined as the absolute difference between the test-retest scans divided by the mean of test-retest. Geodesic Information Flows [10] was used to obtain a white matter (WM) segmentation from the 3D-T1 and all ICVF values were investigated only within the WM.Example ICVF maps are shown in Fig. 1. A clear difference can be observed between different methods, with NODDI notably giving a larger ICVF estimate than the other three methods. Regional variations are also clear, with particularly the splenium of the corpus callosum showing, on average, a higher ICVF than the rest of the brain. Quantification of ICVF between single and crossing fibre voxel shows very different patterns between methods (Fig. 2 and Table 1). CHARMED and DKI show distinctly different histogram modes and medians for ICVF between one, two, and three-fibre voxels whereas NODDI shows no variation with fibre complexity. Fig. 3 and Table 2 show the repeatability of ICVF estimation measured by the coefficient of variation (CV). Here, CHARMED stand out in having a markedly higher CV than NODDI and DKI.There is pronounced variation in intracellular volume fraction estimates between available methods. Most notably, NODDI estimates ICVF values to be 30-40% larger than CHARMED and DKI-derived values. Determining which of these methods reflects the in vivo ICVF most accurately is difficult, as no ground truth exists. Simulations can provide some indications, e.g., Campbell et al [11] found that NODDI ICVF estimates are comparable between single and crossing fibre voxels but such simulations are limited to rather straightforward crossings [12] that do not necessarily reflect the entirety of complex fibre configurations. In the absence of a ground truth, reproducibility of the ICVF estimation is critical, and NODDI and DKI seem to provide superior reproducibility over CHARMED. Interestingly, both CHARMED and DKI demonstrate a difference in ICVF values between 1, 2, and 3-fibre voxels whereas NODDI does not. Given that NODDI is in essence a single-fibre model whereas CHARMED and DKI are not, this difference between the methods might be attributable to the ability to incorporate crossings. Whether this ICVF-dependence on fibre complexity reflects the microstructure is unknown, but one could imagine that packing efficiency is lower for more complex configurations as observed in CHARMED and DKI. Use of these different methods for g-ratio calculations are therefore likely to vary significantly, and care must be taken when comparing ICVF or g-ratio results across studies.