Efficient Evolving Deep Ensemble Medical Image Captioning Network

Abstract

With the advancement in artificial intelligence (AI) based E-healthcare applications, the role of automated diagnosis of various diseases has increased at a rapid rate. However, most of the existing diagnosis models provide results in a binary fashion such as whether the patient is infected with a specific disease or not. But there are many cases where it is required to provide suitable explanatory information such as the patient being infected from a particular disease along with the infection rate. Therefore, in this paper, to provide explanatory information to the doctors and patients, an efficient deep ensemble medical image captioning network (DCNet) is proposed. DCNet ensembles three well-known pre-trained models such as VGG16, ResNet152V2, and DenseNet201. Ensembling of these models achieves better results by preventing an over-fitting problem. However, DCNet is sensitive to its control parameters. Thus, to tune the control parameters, an evolving DCNet (EDC-Net) was proposed. Evolution process is achieved using the self-adaptive parameter control-based differential evolution (SAPCDE). Experimental results show that EDC-Net can efficiently extract the potential features of biomedical images. Comparative analysis shows that on the Open-i dataset, EDC-Net outperforms the existing models in terms of BLUE-1, BLUE-2, BLUE-3, BLUE-4, and kappa statistics (KS) by 1.258%, 1.185%, 1.289%, 1.098%, and 1.548%, respectively.