Brain Blood Vessel Segmentation in Hyperspectral Images Through Linear Operators

Tissue classification tasks that rely on multidimensional data, such as spectral information, sometimes face issues related to the nature of their own characteristics when different biological components share similar spectrum. In a situation of in-vivo brain tumor location during a surgical operation, especially when applying machine learning techniques, relying solely on the spectral information of each sample may not be enough to provide a correct identification of all the tissues involved. In order to overcome this problem, in this work we propose to reduce conflicting classification pixels, i.e. vascular versus tumor tissues. To do so, morphological operators can supply support to a pixel-wise classification by exploiting the spatial characteristics present in vascular tissue. Hence, we have evaluated the suitability of linear operators for brain vessels segmentation in a context of hyperspectral video classification. The parameters of the operator along with the selection of the most suitable spectral band to process were chosen via optimization of the amount of vascular pixels detected and error metrics. The segmentation algorithm was implemented for both CPU and GPU platforms achieving a performance compatible with real-time classification purposes on the last one. Objective results show an average segmentation of the 68% of the vein and arteries present in the ground truth with less than a 10% of error selecting pixels from other tissues of interest such as healthy brain and tumor.