Development of an anomaly detection system for Gibbs artifact identification in amyloid PET imaging.

Abstract

The PET Imaging Site Qualification Program for amyloid positron emission tomography (PET) in Japan includes visual evaluation of the cylinder phantom. This visual evaluation requires observation of the entire image of the phantom and confirmation of the absence of apparent artifacts. Because the evaluation is visually performed, inter-observer differences might exist among evaluators for difficult cases. Therefore, the workload of the staff who perform approval tasks must be reduced, and objective evaluation methods are needed. Thus, we attempted to develop an artificial-intelligence-based objective method for anomaly detection. Three artificial intelligence methods for anomaly detection were developed, and their accuracy was evaluated using AutoEncoder, AnoGAN, and a method combining feature extraction using AlexNet and a one-class support vector machine. In total, 10,207 normal images from 128 facilities and 594 abnormal images from eight facilities, all of which were submitted as part of application for amyloid PET certification, were used. Group five-fold cross-validation was employed for artificial intelligence training and evaluation. In addition, the performance of each artificial intelligence method was assessed using receiver operating characteristic analysis. The areas under the curve for anomaly detection using AutoEncoder, AnoGAN, and the method combining feature extraction using AlexNet and a one-class support vector machine were 0.80 ± 0.04, 0.77 ± 0.03, and 0.99 ± 0.01, respectively. Artificial intelligence effectively distinguished between normal and abnormal images with high accuracy. In the future, its practical implementation is anticipated to reduce the workload in the approval work for the Japanese site qualification program for amyloid PET.