Unsupervised anomaly detection for longitudinal comparison in whole-body PET/CT images.
Authors
Affiliations (5)
Affiliations (5)
- Department of Computational Diagnostic Radiology and Preventive Medicine, The University of Tokyo Hospital, 7-3-1 Hongo, Bunkyo-Ku, Tokyo, Japan. [email protected].
- Department of Radiology, The University of Tokyo Hospital, Tokyo, Japan.
- Division of Radiology and Biomedical Engineering, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan.
- Department of Computational Diagnostic Radiology and Preventive Medicine, The University of Tokyo Hospital, 7-3-1 Hongo, Bunkyo-Ku, Tokyo, Japan.
- Center for Frontier Medical Engineering, Chiba University, Chiba, Japan.
Abstract
This study investigates the utility of unsupervised anomaly detection for longitudinal comparison of whole-body <sup>18</sup>F-fluorodeoxyglucose (FDG)-PET/CT, which (1) reduces false-positive findings compared with subtraction-based methods and (2) enables highlighting newly appearing lesions across diverse anatomical regions without requiring lesion-annotated datasets, unlike supervised approaches. A 3D U-Net was trained exclusively on normal PET/CT images to predict voxel-wise probability distribution parameters of the current PET image conditioned on prior PET, prior CT, and current CT images. During anomaly detection, voxel-wise Z-scores were computed from the predicted parameters and used as an abnormality score map. The performance of the proposed method was compared with that of the subtraction method using slice-wise receiver operating characteristic (ROC) analysis and free-response ROC (FROC) analysis. The area under the slice-wise ROC curve (AUC) was 0.881 for the proposed method and 0.757 for the subtraction method. Region-wise subanalysis demonstrated AUCs of 0.925, 0.918, and 0.824 in the neck, chest, and abdominal regions, respectively, all of which exceeded the corresponding AUCs obtained with the subtraction method. In the FROC analysis, sensitivity at 5.0 false positives per case was 0.728 for the proposed method, compared with 0.375 for the subtraction method. The proposed method effectively suppressed false-positive findings compared with subtraction imaging and successfully highlighted lesions across diverse anatomical regions without requiring lesion-annotated datasets, as opposed to supervised learning approaches.