Tags
#Machine LearningBRISKNet
Dynamic contrast-enhanced (DCE) MRI plays a central role in breast cancer diagnosis and staging by measuring how contrast agents enter and wash out of tissue. These contrast kinetics reflect vascular architecture and tumor biology, and emerging evidence suggests that contrast kinetics within the first two seconds after contrast injection encode previously inaccessible physiological biomarkers relevant to cancer risk and treatment response prediction. However, current breast DCE-MRI acquisition methods are fundamentally limited by a trade-off between spatial and temporal resolution, leaving the early sub-2-second window of contrast dynamics unresolved.
BRISKNet is a self-supervised, physics-informed framework for reconstructing highly undersampled breast DCE-MRI with improved spatiotemporal resolution. The framework integrates principles from compressed sensing, parallel imaging, and non-Cartesian sampling with deep learning-based MRI reconstruction. By avoiding reliance on fully sampled ground-truth data, which is infeasible to obtain in DCE-MRI due to frame rates that inherently require undersampling, BRISKNet enables robust reconstruction at high temporal resolutions. This framework achieves reconstruction quality comparable to iterative methods while enabling faster reconstruction and improved visualization of rapid contrast kinetics. This work provides a foundation for discovering new kinetic biomarkers for personalized breast cancer risk stratification.
Funding and Acknowledgements
This work is partially supported by the University of Chicago’s AI+Science Research Initiative through the Margot and Tom Pritzker Foundation.