Early detection of dementia using artificial intelligence and multimodal features with a focus on neuroimaging: A systematic literature review

Purpose This paper is a systematic literature review of the use of artificial intelligence techniques to detect early dementia. It focuses on multi-modal feature analysis in combination with neuroimaging. The paper examines what past research suggests about issues in the field, what dementia types researchers focus on, what are state-of-the-art methods in the different dementia detection groups, what combinations of modalities (images, text, speech, etc.) are frequently used, how models are evaluated and validated, what datasets researchers use, what are common pre-processing and feature extraction from neuroimages techniques, what are key issues in this research area, and what are potential future research areas. Materials and methods The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) method was used to collect and summarize research in the scope of the defined problem. This study investigated early dementia detection problem from a multi-modal perspective, with neuroimaging being used as one of the modalities. Results Five databases were queried and 2881 sources were identified and processed in the literature review. 59 sources were selected after eligibility assessment. The study identified all points defined in the purpose of the research. Conclusions The main findings of the study were that Alzheimer’s disease and Mild Cognitive Impairment (MCI) are the most researched dementia types in the field; typical choice for dementia detection is Machine Learning (ML) methods; the most popular modalities combination is T1w + Fluorodeoxyglucose - Positron Emission Tomography (FDG-PET); accuracy, sensitivity and specificity are the main evaluation metrics used by the researchers; k-fold validation is being used the most; Alzheimer’s disease neuroimaging initiative (ADNI) is the most used dataset by researchers; intensity and spacial normalization, skull stripping and segmentation are the most common pre-processing techniques for neuroimages; voxel average intensities are being used the most as features in classification extracted from neuroimages; explainability still persists as one of the main issues in adoption of developed methods in clinical practise; there is a lack of studies on Vascular dementia, Frontotemporal dementia, Parkinson’s disease and Huntington’s disease.