Abstract : Fluorescence optical imaging locates biological targets such as tumors tagged with fluorescent markers. For medical diagnostic applications where thickness of media reaches a few centimeters, unwanted autofluorescence of tissues prevents the detection of fluorescence signal of interest : its removal is a sine qua non condition to an accurate tumor localization. The aims of this thesis was to spectrally study natural fluorescence of tissues and to develop a blind source separation method to remove autofluorescence contribution from measurements. Nonnegative Matrix Factorization method was chosen, original algorithms were proposed, and tested on real data. We proved our method is highly efficient to improve the detection of markers and the localization of tumors in Fluorescence Diffuse Optical Tomography reconstructions.