Parameterising the SA-UNet using a Genetic Algorithm

Deep learning is an excellent way for effectively addressing image processing, and several Neural Networks designs have been explored in this area. The Spatial Attention U-Net architecture, a version of the famous U-Net but which uses DropBlock and an attention block as well as the common U-Net convolutional blocks, is one notable example. Finding the best combination of hyper-parameters is expensive, time consuming and needs expert input. We show the genetic algorithm can be utilized to automatically determine the optimal combination of Spatial Attention U-Net hyper-parameters to train a model to solve a Retinal Blood Vessel Segmentation problem. Our new approach is able to find a model with an accuracy measure of 0.9855, an improvement from our previous experimentation which found a model with accuracy measure of 0.9751. Our new methods exhibit competitive performance with other state-of-the-art Retinal Blood Vessel Segmentation techniques.