Simulating dynamic tumor contrast enhancement in breast MRI using conditional generative adversarial networks.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Medical Image Synthesis
• Breast Cancer Diagnostics

Background:

• Deep generative models and synthetic data are crucial for advancing computer-assisted diagnosis and treatment.
• Virtual contrast enhancement in breast MRI offers a promising alternative to physical contrast agent injection.
• Physical contrast agents are invasive, costly, and sometimes contraindicated, limiting diagnostic capabilities.

Purpose of the Study:

• To explore the generation of virtual contrast enhancement in breast MRI using deep generative models.
• To enable lesion localization and categorization without physical contrast agent injection.
• To assist patients for whom contrast agent injection is contraindicated.

Main Methods:

• Proposed a framework for synthetic data properties, introducing the scaled aggregate measure (SAMe) for evaluation.
• Utilized a conditional generative adversarial network to translate non-contrast-enhanced T1-weighted MRI to DCE-MRI.
• Extended the model to jointly generate multiple DCE-MRI time points and investigated 3D U-Net segmentation with synthetic data.

Main Results:

• The SAMe metric demonstrated value in evaluating generative models.
• Virtual contrast injection showed potential for accurate tumor detection and localization.
• Segmentation models augmented with synthetic data exhibited increased robustness to domain shifts.

Conclusions:

• Virtual contrast injection can produce accurate synthetic DCE-MRI images, enhancing breast cancer diagnosis.
• Detecting, localizing, and segmenting tumors using synthetic DCE-MRI is feasible and promising, especially for high-risk patients.
• Joint generation of multiple DCE-MRI sequences improves image quality and enables assessment of tumor response for personalized treatment.