Hybrid U-Net segmentation of vessels in fundus eye images

We present a novel multi-stage method for colour image segmentation, with a primary focus on vessels segmentation in retinal fundus images, using a U-Net based architecture.Our approach tackles challenges posed by varying image resolutions through a coarse-to-fine segmentation pipeline.It begins with a rough segmentation at varying scales, guided by a traditional CNN and progressively refines results to find a target resolution.It culminates with detailed segmentation at a target scale with a smaller window sliding step, compared to previous stages.We train and validate our method using four publicly available datasets FIVES, DRHAGIS, HRF, and STARE -and demonstrate superior performance compared to traditional sliding window techniques.Notably, our model achieves high accuracy with relatively few training images.The entire framework is opensourced and adaptable to a wide range of image segmentation tasks.