Anomaly Detection in PET Images Using Self-Attention PatchCore with Normalizing Flow

Abstract

This study aims to develop an anomaly detection method for cancer in positron emission tomography (PET) images. Our database consists of 1,360 maximum-intensity projection (MIP) images generated from whole-body PET volumes, including 1,030 normal cases and 330 abnormal cases. A self-attention mechanism and normalizing flow are introduced into PatchCore for anomaly detection. The self-attention mechanism is used to consider the feature vectors across the entire image, whereas the normalizing flow transforms the distributions of the feature vectors in the feature maps to approximate the standard normal distribution. Each MIP image is classified as abnormal or normal based on an anomaly score, which is defined as the Euclidean distance to the nearest feature vector in the memory bank of PatchCore. With the proposed method, the classification accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve were 73%, 71%, 75%, and 0.77, respectively, showing improvements compared with the conventional PatchCore (67%, 63%, 70%, and 0.72, respectively). The proposed method could be useful for identifying cancerous lesions and reducing the interpretation time of PET screening.