RESUMEN
3D polarized light imaging (3D-PLI) has been shown to measure the orientation of nerve fibers in post mortem human brains at ultra high resolution. The 3D orientation in each voxel is obtained as a pair of angles, the direction angle and the inclination angle with unknown sign. The sign ambiguity is a major problem for the correct interpretation of fiber orientation. Measurements from a tiltable specimen stage, that are highly sensitive to noise, extract information, which allows drawing conclusions about the true inclination sign. In order to reduce noise, we propose a global classification of the inclination sign, which combines measurements with spatial coherence constraints. The problem is formulated as a second order Markov random field and solved efficiently with graph cuts. We evaluate our approach on synthetic and human brain data. The results of global optimization are compared to independent pixel classification with subsequent edge-preserving smoothing.