Grammar-Guided Evolution of the U-Net

Deep learning is an effective and efficient method for image segmentation. Several neural network designs have been investigated, a notable example being the U-Net which has outperformed other segmentation models in different challenges. The Spatial Attention U-Net is a variation of the U-Net, which utilizes DropBlock and an attention block in addition to the typical U-Net convolutional blocks, which boosted the accuracy of the U-Net and reduced over-fitting. Optimising neural networks is costly, time-consuming and often requires expert guidance to determine the best mix of hyper-parameters for a particular problem. We demonstrate that grammatical evolution (GE) can be used to create U-Net and Spatial Attention U-Net architectures and optimise its choice of hyper-parameters. Our results show improved performance over state-of-the-art models on the Retinal Blood Vessel problem, increasing both AUC, from 0.978 to 0.979 and Accuracy, from 0.964 0.966, from the base models. Crucially, GE can achieve these improvements while finding a model which is 10 times smaller than the base models. A smaller model would enable its use in smart devices, such as smart phones, or in edge computing.