Alzheimer’s disease is the most common cause of dementia in the elderly population (aged 65 years and over), followed by vascular dementia, Lewy body dementia, and rare types of neurodegenerative diseases, including frontotemporal dementia. There is an unmet need to improve diagnosis and prognosis for patients with dementia, as cycles of misdiagnosis and diagnostic delays are challenging scenarios in neurodegenerative diseases. Neuroimaging is routinely used in clinical practice to support the diagnosis of neurodegenerative diseases. Clinical neuroimaging is amenable to errors owing to varying human judgement as the imaging data are complex and multidimensional. Artificial intelligence algorithms (machine learning and deep learning) enable automation of neuroimaging interpretation and may reduce potential bias and ameliorate clinical decision-making. Graph convolutional network-based frameworks implicitly provide multimodal sparse interpretability to support the detection of Alzheimer’s disease and its prodromal stage, mild cognitive impairment. In patients with amyloid-related imaging abnormalities, radiologists had significantly better detection performances with both ARIA-E (sensitivity higher in the assisted/deep learning method [87%] compared to unassisted [71%]) and for ARIA-H signs (sensitivity was higher in assisted [79%] compared to unassisted [69%]). A convolutional neural network method was developed, and external validation predicted final clinical diagnoses of Alzheimer’s disease, dementia with Lewy bodies, mild cognitive impairment due to Alzheimer’s disease, or cognitively normal with FDG-PET. The translation of artificial intelligence to clinical practice is plagued with technical, disease-related, and institutional challenges. The implementation of artificial intelligence methods in clinical practice has the potential to transform the diagnostic and treatment landscape and improve patient health and outcomes.
