Advanced Hybridization and Optimization of DNNs for Medical Imaging: A Survey on Disease Detection Techniques

Abstract

Due to the high classification accuracy and fast computational speed offered by Deep Neural Networks (DNNs), they have been widely used for the design and development of automated Artificial Intelligence (AI) tools for the detection of various diseases. These tools, which are intensive computational learning models, hold tremendous significance in healthcare for identifying various diseases. The primary goal of this review is to understand the applicability and methodology for implementing DNNs, including computational costs, for the classification of distinct diseases from disparate medical imaging datasets. This study presents an extensive survey of DNNs along with their various hybridization forms. To achieve this, the research papers surveyed have been grouped into five categories: pretrained DNNs, hyperparameter-tuned optimized DNNs, hybrid DNNs and ML classifiers, hybrid models with optimization techniques, and meta-heuristics based feature selection DNNs. The major part of this review highlights the significant role of nature-inspired meta-heuristic techniques used for hyperparameter optimization or feature selection algorithms of DNNs. Besides the frameworks and computational costs, descriptions of disparate medical image datasets and image preprocessing techniques have also been discussed under each category. Furthermore, a comparative analysis for each category has been performed on the basis of different parameters, including the type and size of datasets used, image preprocessing, methodology (as per the mentioned category), and performance (in terms of classification accuracy). This study also presents a bibliometric analysis based on the publication count of various articles related to hyperparameter-tuned optimized DNNs and meta-heuristic based feature selection DNNs. This review aims to assist potential AI researchers in choosing the most sound and appropriate DNN-based techniques for disease detection and prediction, all consolidated into a one single research paper.