Generative artificial intelligence enables the generation of bone scintigraphy images and improves generalization of deep learning models in data-constrained environments

• DOI: 10.1007/s00259-025-07091-8
• Publication Date: 2025-01-29T05:50:04Z
• Authors: David Haberl; Jing Ning; Kilian Kluge; Katarina Kumpf; Josef Yu; Zewen Jiang; Claudia Constantino; Alice Monaci; Maria Starace; Alexander R. Haug; Raffaella Calabretta; Luca Camoni; Francesco Bertagna; Katharina Mascher...; Felix Hofer; Domenico Albano; Roberto Sciagra; Francisco Oliveira; Durval Costa; Christian Nitsche; Marcus Hacker; Clemens P. Spielv...
• Keywords: Generative model; Bone scintigraphy

Abstract Purpose Advancements of deep learning in medical imaging are often constrained by the limited availability large, annotated datasets, resulting underperforming models when deployed under real-world conditions. This study investigated a generative artificial intelligence (AI) approach to create synthetic images taking example bone scintigraphy scans, increase data diversity small-scale datasets for more effective model training and improved generalization. Methods We trained on 99m Tc-bone scans from 9,170 patients one center generate high-quality fully anonymized representing two distinct disease patterns: abnormal uptake indicative (i) metastases (ii) cardiac amyloidosis. A blinded reader was performed assess clinical validity quality generated data. added value augmenting an independent small single-center dataset with detect downstream classification task. tested this 7,472 6,448 across four external sites cross-tracer cross-scanner setting associated predictions outcomes. Results The high were confirmed readers, who unable distinguish real (average accuracy: 0.48% [95% CI 0.46–0.51]), disagreeing 239 (60%) 400 cases (Fleiss’ kappa: 0.18). Adding set performance mean (± SD) 33(± 10)% AUC ( p < 0.0001) detecting 5(± 4)% amyloidosis both internal testing cohorts, compared without Patients predicted had adverse outcomes (log-rank: 0.0001). Conclusions Generative AI enables targeted generation different Our findings point potential overcome challenges sharing developing reliable prognostic data-limited environments. Graphical abstract