In ultra-low concentration suspensions, particle number fluctuations in the scattering volume add a long delay component to the intensity autocorrelation function (ACF) in dynamic light scattering (DLS) measurements. This gives a strong artifact peak in the particle size distribution (PSD) recovered. To improve the accuracy of DLS at ultra-low concentrations, we analyzed the different decay characteristics of particle Brownian motion and particle number fluctuation in the ACF. By differentiating the ACF we were able to identify and separate the number fluctuation term and then analyze the ACF to recover the PSD. The results for simulated DLS data at 151nm and 690nm diameters with average particle numbers of 6, 12, 24 and 48 in the scattering volume at four noise levels show that, compared with the usual DLS data processing method, inversion of the ACF after the separation of the number fluctuation term effectively eliminates the strong artifact peaks, and the relative errors of peak positions and distribution errors are significantly reduced. This was further verified with experimental results from samples of standard polystyrene spheres.