Cascaded coarse-to-fine network with hybrid loss for eyeball segmentation in CT image

Eyeball segmentation in computed tomography (CT) images is the basis of computer-assisted diagnosis and surgery navigation system for eye diseases. Fully automatic eyeball segmentation is challenging due to the blurry boundaries, low contrast, and small proportion the eyeball occupied. We propose a framework based on cascaded coarse-to-fine network, combined with hybrid loss function for eyeball segmentation in CT images. The application of refinement module optimizes the coarse segmentation result. The hybrid loss function composed of CE, IoU, and SSIM simultaneously supervises the region and boundaries of the segmentation. Experiments on 4590 2D head CT images show that our method can effectively maintain the eyeball structure with clear boundaries and reduce the false-positive prediction in noneyeball regions.