Images Segmentation using Deep Learning Algorithms and Metaheuristics

Deep learning is a subset of machine learning that encompasses a variety of neural network architectures used to perform diverse computer vision tasks such as medical image classification and segmentation, which are time-consuming, effortful, delicate, and extremely tedious for doctors. The high variability of shape, location, size and texture of the medical images as well as the noise and parasites that degrade the image quality present a big problem for the segmentation process, therefore, a various segmentation methods based on deep learning have been proposed in the literature to fully automated the segmentation process. At the same time, the large number of hyperparameters of a deep learning algorithm in general and of a convolutional neural network in particular presents a problem when developing an automatic segmentation system with an appropriate structure and hyperparameters. Metaheuristics are approximate optimization methods to solve this type of problems. In this study, we review the most used and efficient segmentation methods based on deep learning for medical images segmentation, their optimization with metaheuristics as well as we compared three deep CNN encoder-decoder architectures, namely FCN, SegNet and Unet. These architectures trained and tested on MRI (Magnetic resonance imaging) images in order to study each of those architectures, compare them and finally choose the most efficient model.