Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Harmonizing PET Imaging Across Platforms: A Deep Learning Solution for Neurodegenerative Disease
A new vision transformer framework eliminates cross-platform PET imaging biases, enabling consistent diagnostic standards across different scanner types and manufacturers—a critical breakthrough for neurodegenerative disease assessment.
Background
PET imaging is critical for diagnosing neurodegenerative diseases. The problem? Quantification varies significantly between PET-MRI and PET-CT scanners from different manufacturers. This inconsistency makes reliable cross-platform comparisons nearly impossible and throws a wrench into multicenter clinical trials.
Key Findings
Wang, Zhong, Xu and colleagues developed a vision transformer autoencoder that aligns PET-MRI imaging to PET-CT standards using contrastive learning and attention-guided residual correction. The results speak for themselves:
- Cross-platform bias reduced by >80% in paired same-day scans
- Zero-shot generalization to unseen tracers without retraining
- Amyloid Centiloid discrepancies dropped from 23.6 to 4.1 in multicenter validation (420 patients, three sites, four vendors)
- Tau SUVR thresholds successfully aligned across platforms
Why It Matters
What’s the real impact? This framework enables consistent diagnostic thresholds across platforms and reliable longitudinal monitoring when patients switch between scanner types. It makes multicenter trials more efficient, smooths out therapeutic workflows, and reduces radiation exposure through PET-MRI preference.
Limitations
Validation happened on paired same-day scans. Whether this approach holds up across different timepoints in real-world settings—that’s still an open question.
Original paper: A unified deep learning framework for cross-platform harmonization of multi-tracer PET quantification in neurodegenerative disease. — NPJ digital medicine. 10.1038/s41746-026-02570-0
