Research ArticleClinical Investigation

Facial Anonymization and Privacy Concerns in Total-Body PET/CT

Aaron R. Selfridge, Benjamin A. Spencer, Yasser G. Abdelhafez, Keisuke Nakagawa, John D. Tupin and Ramsey D. Badawi
Journal of Nuclear Medicine August 2023, 64 (8) 1304-1309; DOI: https://doi.org/10.2967/jnumed.122.265280

Article Figures & Data

Figures

  • Figure
  • FIGURE 1.
    FIGURE 1.

    (Top) Faces rendered from CT images at 6 time points. Ninety-minute time point used clinical low-dose CT protocol. All others used research ultra-low-dose protocol. (Bottom) Rendering of defaced images at each time point.

  • FIGURE 2.
    FIGURE 2.

    (Top) Faces rendered from PET images at 6 time points. (Bottom) Rendering of defaced images at each time point.

  • FIGURE 3.
    FIGURE 3.

    (A) Facial embeddings for 15 full-dose cohort subjects at 6 time points, plotted in 2 dimensions using t-SNE. Before defacing, facial embeddings are highly clustered. (B) After defacing, data are no longer clustered.

  • FIGURE 4.
    FIGURE 4.

    (A) Facial embeddings for all 30 full- and low-dose cohort subjects at first 3 time points. Before defacing, clusters are present for each participant. (B) After defacing, participants are not uniquely associated with clusters.

  • FIGURE 5.
    FIGURE 5.

    (A) Facial embeddings for 15 full-dose cohort subjects at 6 time points. Clustered facial embeddings for PET images indicate modest but significant (P < 0.05) degree of identifiability. (B) No clusters are present after defacing. Within-cluster deviation is significantly greater than between-cluster deviation.

  • FIGURE 6.
    FIGURE 6.

    (A) Grid showing CT slices before and after defacing. PET images are reconstructed from same raw data, with different CT scans for attenuation correction. (B) Normalized difference image showing percentage change in PET activity. High-intensity region in brain largely overlaps ventricles, which have low [18F]FDG uptake. (C) Absolute difference in PET SUV.

  • FIGURE 7.
    FIGURE 7.

    (A) Percentage change when using defaced CT for PET attenuation correction. Values were measured at 5 spheric ROIs along cerebral cortex. Error bars correspond to SD within ROI. (B) Difference image and corresponding CT slice, overlaid with 5 ROIs.

Tables

  • TABLE 1.

    Scan Details for 30 Subjects Imaged at 3 or 6 Time Points

    CohortParticipants (n)ModalityInitial*90 min3 h6 h9 h12 h
    Low-dose cohort15CT parameter140 kVp/ 5 mAs140 kVp/ 50 mAs140 kVp/ 5 mAsNANANA
    PET activity†15 MBq11 MBq6.3 MBqNANANA
    Full-dose cohort15CT parameter140 kVp/ 5 mAs140 kVp/ 50 mAs140 kVp/ 5 mAs140 kVp/ 5 mAs140 kVp/ 5 mAs140 kVp/ 5 mAs
    PET activity†288 MBq209 MBq117 MBq36.8 MBq11.6 MBq3.65 MBq

    • ↵* Initial CT and PET images were at 0 and 40 min, respectively.

    • ↵† Reported PET activity is actual injected dose averaged over all subjects. Later time points indicate remaining activity after decay of [18F] (not accounting for excreted activity).

    • NA = not applicable.

  • TABLE 2.

    Identifiability of PET Images Using Classifier Trained with PET (PET-to-PET) or CT (PET-to-CT) Images

    Identifiability (%)Initial90 min3 h6 h9 h12 h
    PET-to-PET64.350.057.157.128.67.1
    PET-to-CT50.042.928.621.428.614.3

Additional Files

Back to top

In this issue

Print
Download PDF
Article Alerts
Email Article
Citation Tools
Bookmark this article

Jump to section

Keywords