Multi-frequency electrical impedance tomography (MFEIT) reconstructs the image of conductivity inside the human body based on the dependence of tissue conductivity on frequency. As there exist differences in the conductivity over frequency between blood, ischemic cortical tissue and normal cortical tissue, MFEIT has potential application in the detection of acute stroke. However, because the conductivity distribution of the human head is highly inhomogeneous and the conductivities of normal head tissue and stroke lesion tissue both change with frequency, the anomaly and normal head tissues are often mixed together in the reconstructed image, which makes it difficult to discern the anomaly. Here we present a spectral decomposition frequency-difference (SD-FD) imaging algorithm in an attempt to address this issue: firstly, we reconstruct so-called EIT spectral images according to the conductivity spectra of tissues; secondly, we obtain the EIT image of the anomaly from the spectral images by using independent component analysis. The results show that the proposed algorithm is capable of detecting the anomaly in a numerical head phantom, as well as in a realistic human head tank with frequency-dependent and heterogeneous conductivities distribution. The proposed SD-FD algorithm may support MFEIT use for human stroke imaging in the future.