Evolutionary Optimization of Convolutional Neural Network Architecture Design for Thoracic X-Ray Image Classification

Abstract

Chest X-Ray images are among the most used tools in medical diagnosis of various hearts and lung abnormalities and infections that could cause pneumonia, severe acute respiratory syndrome, septic shock, failure of multiple organs, and even death. Although such kind of images could be obtained at low cost, the lacking of qualified radiologists limits the exploitation of the X-Ray imaging technology. For these reasons, researchers have proposed the use of deep learning techniques to develop computer-assisted diagnosis systems. Among the most used techniques that have shown great performance in image classification, we find the Convolutional Neural Network (CNN). According to the literature, a good number of CNN architectures already exist. Unfortunately, there are no guidelines to design a specific architecture for a particular task; therefore, such design is still very subjective and mainly depends on the expertise and knowledge of data scientists. Motivated by these observations, we propose in this paper an automated method of CNN design for X-Ray image classification. We demonstrate that the CNN design can be seen as an optimization problem and we propose an Evolutionary algorithm (EA) that evolves a population of CNN architectures, with the aim to output an optimized one. Thanks to the ability of the EA to vary the graphs topologies of convolution blocks, the architecture search space is intelligently sampled approximating the optimal possible CNN architecture. Our proposed evolutionary method is validated by means of a set of comparative experiments with respect to relevant state-of-art architectures coming from three-generation approaches, namely: manual crafting, reinforcement learning-based design, and evolutionary optimization. The obtained results show the merits of our proposal based on the detection of the thoracic anomalies in the X-Ray images.