Breakthrough Applications of Generative Artificial Intelligence in Nuclear Medicine: Exploring the Potential of Synthetic Bone Scintigraphy Images and Their Application in Deep Learning

Background and Research Questions

In recent years, the rapid development of Artificial Intelligence (AI) has revolutionized medical imaging analysis. For instance, deep neural networks have shown great potential in disease diagnosis, anatomical structure segmentation, patient prognosis prediction, and treatment response evaluation. However, the widespread application of these technologies often relies on large-scale, accurately labeled datasets. In the medical field, obtaining such large-scale annotated datasets is both expensive and time-consuming, especially with data sharing strictly restricted due to patient privacy protection. The limited availability of data leads to suboptimal performance and difficulty in generalization of deep learning models in real-life scenarios. This dilemma is particularly evident in distributed research requiring the pooling of data across multiple centers.

On the other hand, the rise of Generative AI provides an innovative solution to the problem of data scarcity. By generating synthetic data, researchers can expand limited datasets and improve model training effectiveness. However, existing generative AI research mainly focuses on conventional imaging domains (such as chest X-rays or brain CT) and remains unexplored in the field of molecular imaging. This paper explores the frontier application of generative AI in synthesizing medical data and optimizing deep learning models, using bone scintigraphy in nuclear medicine as an example.

Paper Overview

The paper, titled “Generative Artificial Intelligence Enables the Generation of Bone Scintigraphy Images and Improves Generalization of Deep Learning Models in Data-Constrained Environments,” is published in the European Journal of Nuclear Medicine and Molecular Imaging and authored by David Haberl and other scholars from the Medical University of Vienna, University of Brescia, University of Florence, Champalimaud Foundation, and West China Hospital of Sichuan University. Officially accepted on January 11, 2025, this is a significant research result in the field integrating generative artificial intelligence and nuclear medicine.

Research Design and Methods

The aim of this study is to develop high-quality synthetic bone scintigraphy images using generative AI techniques to fill existing data gaps and improve the performance and generalization of deep learning classification models. The research primarily includes the following modules:

1. Study Subjects and Datasets

The study utilized datasets from five centers, involving a total of 15,799 patients and 16,823 scans. These data include specific sample sources:

• Vienna General Hospital provided the largest training dataset (9,170 patients, 2010-2020) for the development and training of the generative model.
• ASST Spedali Civili of Brescia: 181 patients, simulating a small-scale single-center dataset.
• Three other external validation centers include:
  1. Careggi University Hospital, Florence (200 patients).
  2. Champalimaud Foundation (674 patients).
  3. West China Hospital of Sichuan University (3,128 patients).

These datasets include two bone scintigraphy tracers labeled with 99mTc (99mTc-DPD and 99mTc-HMDP) and covered two pathological features: Bone Metastases and Cardiac Amyloidosis.

2. Image Generation Model and Optimization

The study used StyleGAN2 (a deep generative adversarial network) to generate bone scintigraphy images. The key feature of the model is the introduction of conditioning variables, allowing it to generate images with specific features based on clinical pathology. The generation process includes the following technical details:

1. Model Training: Input real annotated pathological features (e.g., abnormal tracer distribution), and the model learns to generate high-resolution images (1024×256 pixel resolution) with specific pathological features.
2. Image Screening: Use a Convolutional Neural Network (CNN) to re-verify the generated images, retaining only those that match the pathological features.
3. Data Distribution Verification: Utilize U-MAP to visually compare the generated data with real data to verify whether the synthetic samples faithfully reflect the real data distribution.

3. Data Privacy and Ethical Verification

To ensure patient privacy, a similarity analysis was performed on each generated image to ensure they were not simple replicas of training data. The study received approval from the Ethics Committee of the Medical University of Vienna, with no additional patient record consent required.

4. Model Validation and Classification Tasks

The paper designed three independent research scenarios to verify the practical value of generative data for deep learning models:

• Baseline Scenario: Train the model using only 181 real images from the Brescia center.
• Mixed Data Training: Incorporate synthetic data (with expansion to a 1:50 ratio) to enhance model performance.
• Pure Synthetic Data Scenario: Train the model using entirely synthetic data to verify the efficacy of synthetic data when real data is unavailable.

The study ultimately developed two deep learning models for classification tasks to detect bone metastases and cardiac amyloidosis and validated the models’ generalization ability on four external center datasets.

Main Research Results

1. Quality Evaluation of Synthetic Data

In a blind reading experiment, four nuclear medicine experts could not effectively distinguish between real and synthetic images, with identification accuracy only at the level of random chance (0.48), indicating the high quality of synthetic images. Additionally, similarity analysis ensured the generated images were not direct replicas of training data, showing high independence in pixel-level error and structural similarity metrics.

2. Significant Improvement in Deep Learning Model Performance

• Bone Metastases Detection Task: Incorporating synthetic data into training resulted in an average AUC improvement of 33% compared to using small-scale real data alone.
• Cardiac Amyloidosis Detection Task: The average AUC increased by 5% across centers after introducing synthetic data.

The improvement was particularly notable in early-stage small-scale data scenarios, demonstrating the supplementary value of synthetic data in data-scarce contexts.

3. Prediction and Clinical Outcome Association

Model predictions of abnormal tracer distribution were significantly associated with patient clinical outcomes:

• Patients predicted with bone metastases exhibited significantly increased risk of all-cause mortality (risk ratio exceeding 3 times).
• Patients predicted with cardiac amyloidosis showed significantly increased risk of future heart failure hospitalization (risk ratio over 5 times).

These findings validate the clinical utility of model predictions, further emphasizing the significance of synthetic data.

Research Significance and Highlights

1. Scientific and Application Value

This paper extends the boundaries of generative AI in the field of molecular imaging. By generating synthetic medical imaging data, it not only overcomes barriers of data sharing and privacy protection but also significantly enhances the generalization ability of deep learning models. This research provides an important technical tool for medical imaging research under data-scarce conditions.

2. Methodological Innovation

Key innovations include conditional variable-controlled image generation, CNN-based result verification, and data distribution comparison. These methods ensure the authenticity and diversity of synthetic data and provide a reference framework for future applications of generative AI.

3. Clinical Translation Potential

The association between classification model predictions and adverse patient outcomes underscores the substantial potential of synthetic data in building prognostic models. This technology is particularly significant for small hospitals and rare disease research scenarios.

Summary and Outlook

This study showcases for the first time the practical application potential of generative AI in the bone scintigraphy field. By generating high-quality synthetic data, the study effectively alleviates the bottleneck of data scarcity, proving that synthetic data not only enhances the performance of deep learning models but also possesses clinical practical and analytical value. Although future optimization of generation algorithms and addressing privacy protection remain necessary, this research lays a solid foundation for further integration of generative AI and medical imaging, showing tremendous expansion potential.