A significant advantage of Raman spectroscopy as a noninvasive optical technique is its ability to detect subtle molecular or biochemical signatures within tissue. One of the major challenges for biomedical Raman spectroscopy is the removal of intrinsic autofluorescence background signals, which are usually a few orders of magnitude stronger than those arising from Raman scattering. A number of methods have been proposed for fluorescence background removal including excitation wavelength shifting, Fourier transformation, time gating, and simple or modified polynomial fitting. The single polynomial and the modified multi-polynomial fitting methods are relatively simple and effective, and thus are widely used in biological applications. However, their performance in real-time in vivo applications and low signal-to-noise ratio environments is sub-optimal. An improved automated algorithm for fluorescence removal has been developed based on modified multi-polynomial fitting, but with the addition of (1) a peak-removal procedure during the first iteration, and (2) a statistical method to account for signal noise effects. Experimental results demonstrate that this approach improves the automated rejection of the fluorescence background during real-time Raman spectroscopy and for in vivo measurements characterized by low signal-to-noise ratios.