In order to reduce both acquisition and reconstruction times, illumination and detection in fluorescence diffuse optical tomography (FDOT) have recently evolved from a point-based to a pattern-based approach. The use of structured illumination, offering the ability to project any pattern of light onto the object, associated with the compression of the acquired fluorescence images has paved the way for a new generation of fast reconstruction algorithms for FDOT. However, the choice of the most appropriate set of source patterns is still an open problem. Here, the use of typical source patterns is investigated on experimental data. Reconstructions of similar qualities are obtained for the different types of source patterns. We found that the performances of structured illumination are limited by the required positivity of the source patterns. To alleviate this problem, we introduce a novel method, namely the virtual source pattern method, which allows for considering any kind of patterns, e.g., with negative and complex intensities. This new method provides a significant increase of the contrast of the reconstruction and also a reduction of the reconstruction error, especially when virtual wavelet source patterns are considered.