Carotid Bifurcation Segmentation Based On The Deep Neural Network UNet

Carotid stenosis is an essential biomarker in clinical practice for risk assessment and treatment planning of stroke.Accurate carotid segmentation is a prerequisite for quantifying the degree of stenosis.However,for the lumen segmentation task in the computed tomography angiography(CTA)image,most of the traditional image segmentation methods can not completely get rid of the dependence on human intervention.Recently,the deep learning method displays its outstanding performance on many tasks in the field of computer vision,making it a promising technique to achieve automatic semantic segmentation of carotid bifurcation.Taking three-dimensional head and neck CTA images as the research objects,this study implemented a fully automatic carotid bifurcation segmentation algorithm using a deep convolutional neural network.Based on UNet,taking advantages of residual connections,dilated convolution and deep supervision strategy,this study built up a new network architecture.Moreover,different combinations of loss functions were explored under the architecture to deal with the class imbalance between foreground and background voxels in the segmentation task.By adopting a two-stage strategy,the research realized the segmentation of tiny objects from massive volumes.With 15 training cases,the performance evaluation was done on 41 testing cases and the proposed method achieved 82.3%dice similarity coefficient.It's the first time that deep learning is utilized on carotid bifurcation segmentation in 3D CTA images,implying that deep learning is a promising solution to the problem of extracting carotid bifurcation lumens fully automatically.