Vessel segmentation with deep learning (Conference Presentation)

Vessel segmentation, which is to distinguish blood vessels from the surrounding tissue in images, is a pre-processing step that is often required for the analysis of a vascular network. Three-dimensional segmentation is often challenging in the presence of noise, and a simple thresholding method usually does not work well. Here we have integrated features extracted from 3D images obtained from two-photon in-vivo microscopy with deep learning to do vessel segmentation. The inputs are eigenvalues of the Hessian matrix for each voxel for three different Gaussian filters of width 2, 3, 4 μm and the intensity normalized within the x-y plane. The network is composed of 3-5 layers and each with 3-6 hidden units and is trained for two mouse brain vasculature networks and tested on a third one. The results show a significant improvement compared to a simple thresholding method and are going to be compared with other segmentation methods such as particle filters and enhancement filters. Preliminary results of segmenting OCT data are also obtained.