A graphics processing unit (GPU) accelerated Monte Carlo (MC) approach is developed for modeling photon migration in an arbitrarily complex turbid medium, where the diffusion equation (DE) might behave an ineffective modeling tool. Then an image reconstruction algorithm of time-domain fluorescence diffuse optical tomography is proposed based on the developed GPU-accelerated MC calculations, within the framework of the generalized pulse spectrum technique. Simulated results show that the MC-based approach retrieves on the position and shape of the targets in complexly structured domain that include low absorbing and high scattering, low absorbing and low scattering, high absorbing and low scattering, high absorbing and high scattering, and/or void regions, with higher accuracy than the DE-based one, demonstrating the improved generality of the proposed method.