Related Experiment Video
Updated: Aug 6, 2026
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Scanner-agnostic artificial intelligence approach for fast bone scintigraphy
Vinicius de Oliveira Menezes1,2, Cleiton Cavalcante Queiroz2,3, Antônio Augusto Silva Oliveira1
1Department of Diagnostic Imaging, Nuclear Medicine Division, Hospital das Clínicas, Federal University of Pernambuco (HC/UFPE-HU Brasil), Recife, Pernambuco, Brazil.
Journal of Applied Clinical Medical Physics
|July 22, 2026
Summary
A novel adaptive-diffusion U-Net enables diagnostic-quality bone scintigraphy from reduced scan times or radiation doses. This scanner-agnostic deep learning approach preserves diagnostic integrity across different gamma cameras without retraining.
Area of Science:
- Medical Imaging
- Nuclear Medicine
- Artificial Intelligence
Background:
- Current bone scintigraphy requires lengthy scans (10-15 min) for optimal image quality.
- Existing deep learning denoisers often need camera-specific retraining, limiting their broad application.
Purpose of the Study:
- To introduce a scanner-agnostic adaptive-diffusion U-Net for reconstructing high-quality bone scintigraphy images.
- To enable diagnostic-grade imaging from reduced acquisition times or radiation doses without scanner-specific tuning.
Main Methods:
- A multi-institutional dataset of 3635 bone scintigraphy studies from four gamma camera models was used.
- The adaptive-diffusion U-Net, combining a U-Net backbone with a diffusion layer, was trained on Poisson-thinned data (10%-70% counts).
- Prospective validation included 60 patients and head-to-head comparisons with standard acquisitions, with blinded physician ratings.
Main Results:
- Deep learning reconstructions significantly improved image quality metrics (SSIM, PSNR) and reduced perceptual errors (LPIPS) at 50% counts (p < 0.001).
- Physician ratings showed comparable diagnostic performance for accelerated deep learning scans versus standard scans (Likert scores 4.4 vs 4.5).
- No diagnostic discrepancies were reported between deep learning-reconstructed and routine clinical images.
Conclusions:
- The adaptive-diffusion U-Net effectively supports reduced-time/lower-dose bone scintigraphy protocols.
- The approach maintains diagnostic integrity across multiple scanner vendors without retraining.
- This facilitates scalable, radiation-sparing adoption in clinical practice.
