Fully Unsupervised Deep Spectral Clustering for Retinal Vessel Segmentation

Blood vessel segmentation plays a crucial role in the diagnosis and treatment of eye diseases. Fully supervised and semi-supervised deep learning-based methods require large, labeled datasets for their training. We propose a novel, fully unsupervised deep learning segmentation framework that implements the spectral clustering algorithm to address this issue. A deep neural network is used to learn the continuous Laplace-Beltrami partial differential operator of the images and solve its eigen decomposition in an end-to-end manner with a tai-lored loss function. The learned eigen decomposition spans the image space and is specifically tuned to the input image. Pixels are mapped to this space, and a k-means clustering algorithm produces the final segmentation results. We validated our method on the publicly available DRIVE and STARE reti-nal vessel datasets, demonstrating qualitative and quantitative results. Our approach outperforms other fully unsupervised methods for this task.