Generative artificial intelligence (AI) enables the creation of large, anonymized image datasets for the development of diagnostic systems. As part of an international collaboration, researchers at MedUni Vienna have developed a generative AI that can generate and process synthetic medical imaging data. The resulting data was then used to develop diagnostic AI systems for multiple diseases. The research team demonstrated that the accuracy of AI models was significantly improved through the use of synthetic imaging data. The results of the study were recently published in the journal “European Journal of Nuclear Medicine and Molecular Imaging”.
The use of AI systems in medical imaging is becoming increasingly important. The accuracy of these systems largely depends on the quality and quantity of the training data. However, clinical datasets are often limited in their usability – whether due to small sample sizes for rare diseases, strict data protection regulations, or the underrepresentation of certain subgroups. In the worst case, this can lead to inaccurate predictions by AI models, especially if the training data is not representative of the overall population.
The generation of synthetic datasets using generative AI offers a promising solution to these challenges. By synthesizing medical imaging data, AI systems can be trained on a broader range of disease patterns without relying on rare or difficult-to-access patient data. This approach was pursued by a research team from the Division of Nuclear Medicine (Department of Biomedical Imaging and Image-guided Therapy) at MedUni Vienna, in a recently published study. A generative AI was trained on more than 9,000 scans from routine clinical examinations in the scintigraphy outpatient department. The model was then used to generate a synthetic image dataset that replicates the characteristic features of real medical imaging data but is entirely newly generated, ensuring that no patient-related information can be traced.
Synthetic Imaging Data of Equal Quality
The quality of the synthetic data was evaluated by four independent physicians. They found no discernible difference between the synthetically generated and real imaging data. The relevance of the synthetic data was further confirmed by an independent research group at the University of Brescia. There, a team of researchers developed an AI system to detect patients suspected of having cardiac amyloidosis or bone metastases, trained using the synthetic imaging data generated in Vienna. The system was subsequently validated using data from more than 6,000 patients from four independent institutions in Europe and Asia. The results showed that integrating synthetic data significantly improved the diagnostic accuracy of the AI system.
“One major advantage of this technology is the protection of patient privacy: since the generated images do not depict real patients, they can be used for research and the development of new AI-assisted diagnostic methods without any data protection risks” summarize David Haberl and Clemens Spielvogel from the MedUni Vienna study team regarding the significance of the study's findings. Moreover, adding synthetically generated data allows to increase data diversity. For example, by specifically incorporating data from underrepresented subgroups, a dataset can be adjusted to improve the accuracy of the resulting AI systems for these subgroups in clinical applications.
Publication: European Journal of Nuclear Medicine and Molecular Imaging
Generative artificial intelligence enables the generation of bone scintigraphy images and improves generalization of deep learning models in data-constrained environments;
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 Mascherbauer, Felix Hofer, Domenico Albano, Roberto Sciagra, Francisco Oliveira, Durval Costa, Christian Nitsche, Marcus Hacker & Clemens P. Spielvogel
https://doi.org/10.1007/s00259-025-07091-8