To detect bone metastases on FDG-PET/CT images with fewer false positives (FPs), we propose a method based on cascade voxel classification by two types of unsupervised AI anomaly detection.Firstly, anomaly detection using Mahalanobis distance from normal bone voxels is adopted at bone voxels to extract voxels with abnormal CT value and SUV coarsely. Secondly, nonlinear anomaly detection using one-class support vector machine (OCSVM) is applied to the coarsely extracted voxels. The OCSVM uses seven local texture features to enhance metastases-like patterns. Finally, local maxima of the enhanced image are detected as the metastasis candidates.An experiment with ten images including 19 metastases showed the result with 89.5 % sensitivity and 59.5 FPs per case. The number of FPs was 93.9 % lower than when OCSVM was used only.This result confirmed the effectiveness of the proposed method.