A Survey on Hybrid Case-Based Reasoning and Deep Learning Systems for Medical Data Classification

Introduction

Artificial intelligence (AI) aims to design and create systems and applications that can imitate human intelligence and assist in the actual solution of complex issues. AI is already making a difference today in its use in all scientific fields and more particularly in the field of medicine. Among the different branches of AI that have received more and more attention from research in medicine, we find case-based reasoning and deep learning. Case-based reasoning (Kolodner & Reasoning, 1993) (Gasmi, Djebbar, & Merouani, 2021) (Chebli, Djebbar, Marouani, & Lounis, 2021) is an approach for solving future issues relying on reusing previously solved issues saved in the case base, where a new issue is resolved using knowledge or similar experiences presented in case form. A case is defined in two parts, where the first is the problem and the second part is the solution of the problem. These cases which are grouped in the base of cases are independent of each other CBR systems are applied for several purposes for example for medical diagnosis (Huang, Chen, & Lee, 2007) (Begum, Ahmed, Funk, Xiong, & Von Schéele, 2009) (Chebli, Djebbar, & Merouani, 2020), medical planning (Cohen, 1997) (Marling & Whitehouse, 2001), classification of diseases (LeBozec, Jaulent, Zapletal, & Degoulet, 1998) (Fan, Chang, Lin, & Hsieh, 2011), etc. The same goes for Deep Learning (DL) (Kwolek, 2005) which has demonstrated remarkable performance in all medical uses including classification, diagnosis, segmentation, and detection of an anomaly. So DL quickly replaced classical AI techniques because it may give a much better explanation of a complicated situation if large datasets are available. Moreover, it can manipulate all forms of medical data, including images, genetic expressions, signals, etc. Although case-based reasoning or deep learning is simple in principle and has been used successfully in medical problem solving, it still suffers from some limitations that may prevent its success. CBR and more precisely the retrieval phase of the CBR cycle which is the essential phase that takes the responsibility of finding the right solution that is then used by the other phases in order to solve the given problem, still remains unclear and invalid in the medical context, and this is due to the set of naive classical methods used in this phase. And to overcome these limitations, several researchers have proposed a hybridization strategy between CBR and DL, with the intention of taking advantage of the properties of each model, and thus obtaining very good performances for the retrieval phase and the other phases of CBR in the medical context. The objective of this research, is to analyze hybrid systems that combine CBR and DL in medicine and other fields, as well as to suggest an architecture to improve the recovery stage of CBR using DL. The rest of the paper is structured as follows: Section 2 presents the background that encompasses a state of the art of CBR, as well as various medical systems that use CBR, as well as developed medical systems that use DL. In Section 3, hybrid medical systems that combine CBR and DL are studied. In Section 4, some results obtained from hybrid systems are discussed, and the CNN architecture and its popular models are presented. Section 5 presents the proposed approach, and in section 6, future research directions are presented. Section 7 presents the conclusions.