Automatic Generation of Optimized Convolutional Neural Networks for Medical Image Classification Using a Genetic Algorithm

Abstract

Convolutional neural networks (CNNs) are at the heart of many state-of-the-art computer-aided detection/diagnosis (CAD) systems for medical imaging. However, there is a trend towards developing task-specific CNN-based models that can reach high-quality results at preliminary design stages but are over-fitted to the training dataset or cannot be deployed in a clinical setting. Also, the design of CAD systems usually focuses on the CNN architecture and omits the proper selection of the network's hyperparameters. We offer a solution to this problem that advances the current state of the generic field of artificial intelligence applied to medical imaging by proposing a genetic algorithm that optimizes the selection of the numerical and categorical network's hyperparameters and, at the same time, automatically generates the CNN models. We obtain competitive standard evaluation metrics compared to the current hybrid and task-specific CNN-based models. Our methodology is validated with the KVASIR dataset showing that our novel framework works for gastrointestinal image classification, obtaining an accuracy of 0.9850, $F_{1}$ value of 0.9400, and a Matthews correlation coefficient of 0.9314.

Note:
Funding Information: This research was possible thanks to funding provided by Consejo Nacional de Ciencia y Tecnolog´ıa (CONACYT).

Conflict of Interests: The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Keywords: deep learning, Hyperparameter optimization, Genetic algorithm, Kvasir dataset