Evaluation of BTXBrain‐Tau, an AI‐powered automated quantification software for Flortaucipir PET images

Abstract

AbstractBackgroundPositron emission tomography (PET) has been used to visualize the distribution of brain biomarkers associated with AD. Flortaucipir (18F‐AV‐1451) is a tau‐specific PET tracer that has been used to detect tau pathology in AD. However, the accurate quantification of Flortaucipir PET images requires an accurate spatial normalization of the image data. We evaluated BTXBrain‐Tau, an AI‐based automated quantification software that does not require an MRI, on the Flortaucipir data. In comparison, performance on automatic quantification was demonstrated against conventional MRI‐based spatial normalization results.MethodIn the previous study, BTXBrain‐Tau was pretrained with approximately 1000 amyloid PET images. To fine‐tune the network, 180 Flortaucipir images were obtained from Seoul National University Hospital. Trained network was evaluated using 146 Flortaucipir images from the ADNI. Each image had been quantified using FreeSurfer at UC Berkeley, and these quantifications served as the “ground truth.” Spatial normalization was also performed using SPM12 with 3D T1 MRI to demonstrate the proposed method’s superiority. The results were compared using four Braak stage‐based composite regions of interest. Statistical analysis of the results was conducted using linear regression and intraclass correlation coefficients.ResultThe linear regression slopes for BTXBrain‐Tau were 0.95, 0.97, 1.03, and 0.95 for four regions of interest (ROIs), outperforming the results of SPM12 (0.81, 0.80, 0.81, 0.79). The R‐squared values also showed superior performance for BTXBrain‐Tau (0.94, 0.98, 0.95 and 0.98) compared to SPM12 (0.88, 0.97, 0.91 and 0.96). ICCs between the ground truth and SPM12 normalization results were 0.93, 0.96, 0.94 and 0.96 for SPM12, however, these values were improved in BTXBrain‐Tau processing (0.97, 0.99, 0.97 and 0.99).ConclusionThe results demonstrated that BTXBrain‐Tau was able to generate spatially normalized images without the use of 3D MRI input. Statistical analysis revealed a high degree of correspondence between the FreeSurfer‐based quantification and the proposed method. Additionally, the study demonstrated the generalizability of the network by assessing its performance on different types of test data, (training on Asia, testing on Caucasian). These results suggest that BTXBrain‐Tau is a promising tool for the accurate quantification of Flortaucipir PET images, which could lead to early detection and precise staging of Alzheimer’s disease.