Supervoxel algorithm for medical image processing

Oversegmentation into supervoxels decreases computational complexity and provides more self-descriptive units than voxels. Most of the existing methods are based on minimization of cost function using different approaches, such as gradient descent. Only a few of them produce 3D segments (supervoxels), which enforces processing of medical data slice-by-slice. We propose a supervoxel method for processing of volumetric medical data. Unlike others, our method incorporates physical distance between voxels. The method iteratively refines initial edges. The cost function contains edge preservation, homogeneity, and regularity terms. Our results are comparable to state-of-the-art and unlike most algorithms, we are able to produce both supervoxels and superpixels. We show that 3D supervoxels provide more statistical information than 2D superpixels with comparable accuracy and that fusion of multiple modalities does not increase accuracy of supervoxels, but improves quality of superpixels.