Scanner-agnostic artificial intelligence approach for fast bone scintigraphy.

PURPOSE: Current bone scintigraphy protocols often demand full-count, 10-15 min scans to preserve image quality, and existing deep-learning (DL) denoisers typically need to be retrained or retuned for each camera manufacturer. We introduce a scanner-agnostic adaptive-diffusion U-Net designed to reconstruct diagnostic-grade images from half-time or half-dose acquisitions without scanner-specific retraining.

METHODS

A multi-institutional retrospective set of 3635 studies from four gamma-camera models was Poisson-thinned to 10%-70% counts and partitioned for training/validation/testing. The network couples a U-Net backbone to a trainable isotropic diffusion layer. Quantitative evaluation (SSIM, PSNR, LPIPS) used 182 test scans. Prospective validation in 60 patients, including 20 studies acquired on an unseen Philips BrightView scanner not represented in the training set, was performed, and images were rated by three blinded nuclear medicine physicians on a 5-point Likert scale. Head-to-head comparisons evaluated the diagnostic fidelity of the DL approach.

RESULTS

At 50% counts, DL reconstructions increased SSIM from 0.903 ± 0.045 to 0.963 ± 0.032 and PSNR from 30.50 ± 4.13 dB to 40.85 ± 4.55 dB compared with noisy images, while reducing LPIPS from 0.05 ± 0.01 to 0.03 ± 0.01 (p < 0.001). Reader studies showed that double-speed DL images achieved Likert scores comparable to standard acquisitions (4.4 ± 0.7 vs 4.5 ± 0.5), with no diagnostic discrepancies reported between DL and routine clinical images.

CONCLUSION

In this multicenter, multi-vendor study, the adaptive-diffusion U-Net effectively supported half-time, and/or lower-activity whole-body bone scintigraphy protocols while preserving diagnostic integrity across scanners and avoiding scanner-specific retraining, supporting scalable, radiation-sparing clinical adoption.