Adaptive Ultrasound Imaging with Physics-Informed NV-Raw2Insights-US AI

Source: HuggingFace Date: 2026-04-28 URL: https://huggingface.co/blog/nvidia/raw2insights-adaptive-ultrasound-imaging

Summary

NVIDIA and Siemens Healthineers released NV-Raw2Insights-US, a physics-informed AI model that processes raw ultrasound channel data rather than reconstructed images. By generating per-patient speed-of-sound maps in real time, it corrects for individual anatomy on the fly — addressing a fundamental limitation of hand-engineered reconstruction pipelines that assume constant tissue properties. Inference runs on Blackwell-class GPUs (IGX Thor or DGX Spark) with raw data streamed via NVIDIA Holoscan Sensor Bridge over DisplayPort.

Implications

• Software-defined medical imaging as a pattern. The architecture — keep existing hardware, replace reconstruction with a GPU inference layer, update continuously via software — mirrors what happened to networking (SDN) and is now reaching radiology equipment. The DisplayPort streaming approach minimizes scanner hardware changes.
• Physics-informed AI as a distinct capability class. Rather than learning purely from labeled data, the model encodes domain physics (acoustics, tissue interaction). This generalizes better under distribution shift and reduces labeled training data requirements — a pattern likely to spread across sensor-heavy verticals (sonar, seismic, NDT).
• Feeds the specialized hardware + edge inference thread: Real-time per-patient adaptation at the scanner requires edge-class GPU compute. DGX Spark at the point of care is the enabling step; this workload justifies that hardware class in clinical settings.