Jie Zhu1, Ashish Raj2, Ramin Zabih3
1Electrical Engineering, Cornell U, Ithaca, NY, USA; 2Radiology, Weill Medical College of Cornell University, New York, NY, USA; 3Computer Science, Cornell U, Ithaca, NY
Segmentation of MR brain images is an important step in many clinical applications, but is challenging due to the vast amount of fine structures. Conventional EMS algorithm fails in presence of noise, partial voluming or fine details and textures. Recent improvement were reported using graph cuts but standard inter-voxel edge weighting in graph cuts fails near small structures. We propose a graph-based combinatorial algorithm called geo-cuts, where image gradient magnitudes and gradient directions determine the weighting, to overcome this problem. Our results show that the overall percentage of correctly classified voxels is higher using the geo-cuts method except when noise is present. The Dice Similarity Measure indicates that both methods work well with geo-cuts generally outperforming standard graph cuts for white matter and CSF and vice versa for gray matter. Visually, geo-cuts gives better performance on both real and synthetic images with respect to fine structures of white matter and CSF.