DenOiS: Dual-Domain Denoising of Observation and Solution in Ultrasound Image Reconstruction
New framework denoises both raw sensor data and final images, enabling training on simulations for real-world use.
A new AI framework called DenOiS, developed by researchers Can Deniz Bezek and Orcun Goksel, tackles a core problem in medical imaging: reconstructing clear pictures from noisy, incomplete data using imperfect physical models. Unlike traditional methods that rely on hand-crafted rules or purely data-driven AI, DenOiS employs a novel dual-domain approach. It first refines the raw sensor observations to correct for noise and model inaccuracies, then uses a diffusion-based plug-and-play (PnP) reconstruction method that remains stable even with missing measurements. This two-stage process allows the system to separate and address errors in both the measurement and solution domains.
The key breakthrough is DenOiS's ability to bridge the simulation-to-reality gap. The model can be trained exclusively on synthetic data yet still produce high-fidelity reconstructions from real, noisy ultrasound scans. The researchers demonstrated this capability for speed-of-sound imaging, a quantitative technique that maps tissue stiffness but is highly sensitive to data quality. By making the reconstruction robust to imperfect real-world conditions, DenOiS could enable more reliable and accessible quantitative ultrasound diagnostics without the need for vast, perfectly labeled clinical datasets.
- Uses a dual-domain approach to denoise both raw sensor data (observations) and the final reconstructed image (solution).
- Employs a diffusion-based plug-and-play (PnP) method that is robust to incomplete or missing measurements.
- Can generalize to real patient data after being trained only on simulations, a major step for practical clinical AI.
Why It Matters
Enables more accurate, quantitative medical imaging from noisy real-world data, potentially improving diagnostic reliability without costly data labeling.