NVIDIA and Siemens Healthineers Launch NV-Raw2Insights-US for Raw Ultrasound Data Processing

NVIDIA, in collaboration with researchers from Siemens Healthineers, has officially introduced NV — Raw2Insights-US, a foundational artificial intelligence model designed to transform diagnostic medical imaging. Published on April 28, 2026, through a Hugging Face blog post, this collaborative effort shifts the paradigm of ultrasound technology by analyzing raw sensor data rather than relying on traditionally reconstructed images. For decades, ultrasound machines have compressed rich sensor measurements into a final image using hand-engineered reconstruction pipelines. This conventional approach fundamentally limits diagnostic potential because it actively discards the intricate details of how sound waves move through human tissue before the clinician even views the screen.

The primary application of this initial Raw2Insights release focuses directly on adaptive image focusing through patient — specific speed — of-sound estimation. Standard clinical ultrasound systems typically operate under a rigid, simplifying assumption that the speed of sound remains entirely constant as it travels throughout the body. In contrast, NV — Raw2Insights-US actively listens to the raw signals captured by the ultrasound probe, calculating how each individual patient uniquely shapes those returning echoes. This capability allows the system to generate a highly personalized sound speed map in a single computational pass.

Implementing this direct — interpretation approach required engineering a solution to a longstanding hardware bottleneck caused by the massive bandwidth of raw ultrasound channel data. Historically, this dense information remained inaccessible within clinical — grade scanners. To extract it, the engineering team utilized the Holoscan Sensor Bridge, an open-source FPGA intellectual property developed by NVIDIA. Operating on an Altera Agilex-7 FPGA development kit, the bridge interfaces with an ACUSON Sequoia ultrasound scanner. The system extracts raw channel data streaming directly from the scanner's DisplayPort outputs, a technique the developers have named Data over DisplayPort.

Once extracted, the Holoscan Sensor Bridge packetizes the dense data and transmits it over Ethernet using Remote Direct Memory Access over Converged Ethernet to an NVIDIA IGX system for data collection. The operational deployment relies heavily on NVIDIA Holoscan, a specialized edge AI sensor processing platform specifically designed for high-performance, real-time medical workloads. Processing occurs on advanced hardware infrastructures, including the NVIDIA IGX Thor and NVIDIA DGX Spark systems. The actual accelerated inference is executed on a Blackwell — class GPU, which rapidly computes the necessary patient — specific sound — speed estimates.

Beyond immediate imaging improvements, the architecture introduces a highly modular foundation for the next generation of artificial intelligence diagnostics. Because the raw channel data is directly loaded into the GPU memory without prior degradation, developers can seamlessly integrate an array of subsequent AI models. The system supports software — only integration, allowing existing medical devices to achieve modern acceleration through the Data over DisplayPort pathway. Furthermore, this software — defined approach to ultrasound ensures continuous capability improvements through routine software updates. To encourage further industry innovation, NVIDIA has made the NV — Raw2Insights-US model weights, associated datasets, and development resources publicly available via GitHub.