Abstract
The application of machine learning (ML) and artificial intelligence (AI) algorithms in medical imaging is an emerging area of interest, particularly in the context of clinical decision-making. Here, we report on the overall performance (i.e., sensitivity, specificity, and accuracy) of commonly used ML/AI techniques including convolutional neural networks (CNNs), support vector machines (SVMs), random forests, and ensemble approaches on the clinically relevant task of distinguishing between Parkinson's disease (PD) participants and matched healthy controls (HC). Our systematic review includes 130 studies from six different imaging modalities - dopamine transporter scans ([123I]Ioflupane single-photon emission computed tomography (SPECT)), positron emission tomography (PET) including [18F]FDG, [18F]DOPA, and [11C]raclopride, structural magnetic resonance imaging (MRI) (T1- and T2-weighted), functional MRI, and diffusion MRI. While some findings were in line with expectations for some modalities, such as the superior performance of dopamine SPECT and PET (> 90% sensitivity, specificity, and accuracy with methods like convolutional neural networks), others were more nuanced with the best-performing class of algorithms depending on the imaging modality, and sometimes, even the data source. Overall, we summarize the emerging trends across studies for each imaging technique and provide valuable recommendations for future lines of inquiry.