For a magnetic induction tomography (MIT) system operating at 10 MHz, the signals produced by a haemorrhagic cerebral stroke were computed using an anatomically realistic, multi-layer, finite element (FE) head model comprising 12 tissues. The eddy-current problem was approximated using the commercial FE package, Comsol Multiphysics, and the numerical techniques employed were validated using a benchmark test. Mesh convergence for the head model was investigated for first- and second-order elements. MIT signals were computed for strokes of different sizes and locations in the brain to judge the sensitivity of the MIT configuration. The results revealed that for a large peripheral stroke (volume 50 cm(3)), 27% of the signals were above the phase noise level achievable in our current data-collection hardware (approximately 20 m degrees). In order to detect the same percentage of the signals due to a centrally located small stroke, a reduction in phase noise to 1 m degrees was necessary.