EEvoU-Net: An ensemble of evolutionary deep fully convolutional neural networks for medical image segmentation

Abstract

Developing a Deep Convolutional Neural Network (DCNN) is a challenging research topic which needs extensive efforts and computation to find a proper network structure and a precise set of hyper-parameters. The problem is that in most of the cases the achieved network works well on the specific application or dataset, and a small to significant changes are required to adapt it for a new one. Besides, the limited number of available labelled images and the required computational infrastructure, make this task even more challenging. Therefore, developing an automatic method that is able to find a network structure and its parameters, while using minimum computation, seems necessary. Evolutionary computation is an optimisation method that can be used to address the mentioned difficulties. This paper proposes an evolutionary-based framework to find a set of precise and small networks for medical image segmentation, and also, an ensemble model to improve the quality of segmentation. To the best of our knowledge, EEvoU-Net is the first ensemble method that utilises a set of evolutionary U–Net–based deep networks for medical image segmentation. In the proposed model, a Genetic Algorithm (GA) is applied to design a set of optimal network structures, along with their parameters, using a new fixed-length encoding strategy to create variable length networks, for medical image segmentation. The proposed model is evaluated using five different, publicly available medical image segmentation datasets. The best found evolutionary networks, outperformed U-Net, ResU-Net, DenseU-Net, NAS U-Net, AdaResU-Net, EvoU-Net, DenseRes, 2D to 3D EvoU-Net, and Attention EvoU-Net in the most of the cases using considerable less trainable parameters. Furthermore, EEvoU-Net as an ensemble of evolutionary networks, has also substantially improved over those previous results.