Adversarial robustness improvement for X-ray bone segmentation using synthetic data created from computed tomography...

Area of Science:

• Medical imaging
• Artificial intelligence
• Deep learning

Background:

• Deep learning in clinical imaging faces challenges with limited annotated data and adversarial susceptibility.
• AI systems for tasks like patient positioning require balanced datasets, which are often unavailable.

Purpose of the Study:

• To generate synthetic X-ray images and annotation masks from 3D CT volumes to create realistic, non-optimally positioned training data.
• To enhance the robustness of AI models trained on imbalanced real-world clinical data.

Main Methods:

• Utilized 3D photon-counting CT volumes to forward project synthetic X-ray images and annotation masks.
• Employed the open-source TotalSegmentator model for annotating clavicles in 3D CT data.
• Evaluated model robustness by comparing real-data-trained models with real-and-synthetic-data-trained models under simulated patient rotation.

Main Results:

• Models trained on both real and synthetic data demonstrated 3% to 15% Dice score improvements across various simulated patient rotation angles compared to real-data-only models.
• Synthetic data effectively augmented underrepresented conditions, leading to increased model robustness.

Conclusions:

• Synthetic data generation is a viable strategy to supplement training datasets in medical AI.
• This approach enhances the robustness of deep learning models for clinical image analysis, particularly for underrepresented scenarios.