Accuracy of deep learning model-assisted amyloid positron emission tomography scan in predicting Alzheimer's disease: A Systematic Review and meta-analysis

Abstract Background and aim Alzheimer's disease (AD) is a neurodegenerative disease that attacks the brain by deposited amyloid-beta and neurofibrillary tangles. This study aimed to evaluate the diagnostic characteristics of deep learning (DL) in predicting AD with an amyloid positron emission tomography (PET) scan. Materials and methods The ISI Web of Science, PubMed, Scopus, Cochrane Library, IEEE Xplore Digital Library, ALZFORUM, CINAHL, Science Direct, PROSPERO, PsycINFO, and EMBASE were systematically searched. We used the revised Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) tool to investigate the quality and potential bias of all studies. The area under the curve (AUC) was calculated for the receiver-operating characteristic (ROC) to determine diagnostic precision. All experiments were viewed with the summary ROC (SROC) plotted as a circle. Results Overall, 17 studies (a total of 15,177 image data) were included in the final meta-analysis following reviewing titles, abstracts, and full papers. The pooled AUC, sensitivity, and specificity of the amyloid tracers were 0.96 (95% CI, 0.78–0.98), 0.92 (95% CI, 0.88–0.95), and 0.90 (95% CI, 0.86–0.93), respectively. The pooled AUC, sensitivity, and specificity of the FDG tracers were 0.95 (95% CI, 0.77–0.96), 0.92 (95% CI, 0.87–0.96), and 0.88 (95% CI, 0.84–0.91), respectively. The pooled AUC, sensitivity, and specificity of the Florbetapir tracers were 0.97 (95% CI, 0.86–0.98), 0.94 (95% CI, 0.87–0.98), and 0.96 (95% CI, 0.87–0.99), respectively. Conclusion Our findings demonstrated that DL provided a more accurate method by amyloid PET images leading to straightforward advances in the early diagnosis of diseases, such as AD.