Research Of Diffusion Tensor Image Segmentation Based On Supervoxel

Diffusion tensor imaging is able to quantify the diffusion properties of the water molecules to probe the microstructural and physiological information of tissues in vivo.Therefore,this powerful technique has been increasingly applied for the clinical applications and scientific research.The wealthy information in the diffusion tensor image can enable accurate segmeantion of a number of important tissues.However,the voxel-wise strategy suffers from the limitation of a high computational complexity due to the large number of voxels and the high dimensionality,which limits their applications for the real-time clinical use.The efficiency problem can potentially be addressed by introducing the supervoxel segmentation technique.The promising technique can group similar voxels into a meaningful supervoxel.The huge number of voxels can be transferred into a small number of supervoxels,which can highly reduce the computational burden.Diffusion tensor image has smooth regions in a local,which makes it suitable for supervoxel based processing.Therefore,this study proposed a novel supervoxel based segmentation framework especially for diffusion tensor image.Considering the features of diffusion tensor image,we presented a novel supervoxel segmentation approach for diffusion tensor images based on tensor invariants and orientation.Firstly,tensor invariant and orientation features were derived to capture the invariant and orientation features of diffusion tensors.Secondly,voxels were classified into uniformly sampled seeds using spatial fuzzy clustering with a novel distance metric by integrating spatial coordinate,invariant and orientation features.Finally,the optimization was iteratively repeated until convergence to generate the final supervoxels.After obtaining supervoxels,this study proposed a novel supervoxel clustering based segmentation method for diffusion tensor image under the spectral clustering based learning framework.The proposed method can learn the discriminative features to better distinguish tissues from background for specific tissue segmentation.At first,geometry and orientation features were extracted from each supervoxel.Secondly,a novel metric learning method was proposed to learn the optimal metric for supervoxel classification.Finally,the optimization model was solved with an alternative strategy to obtain the optimal metric kernel and the labels of supervoxels.The segmentation results were finally achieved from the labels of the supervoxels.The proposed method was evaluated with qualitative and quantative comparisions on the task of corpus callosum segmentation from the brain data.Qualitative evaluation was performed by visualizing two dimensional image and three dimensional surfaces of the segmentation results.Quantative evaluation was performed by calculating commonly utilized metrics,including dice similarity Coefficient(DSC)?Absolute Volume Difference(AVD)?Hausdorff Distance(HD).The proposed method could efficiently achieve accurate segmentation results compared to the traditional methods.