AI-Guided DVT Ultrasound Scanning

Background

Access to advanced diagnostic procedures remains severely limited in rural America, where specialist shortages mean that conditions like deep vein thrombosis (DVT) often go undetected until they become life-threatening. DVT ultrasound scanning requires a skilled practitioner to navigate a probe through precise anatomical positions — a skill that takes years to develop.

This project is part of the ARPA-H PARADIGM program (Platform Accelerating Rural Access to Distributed InteGrated Medical Care), which aims to bring hospital-level care to rural communities through a mobile Care Delivery Platform (CDP). PARADIGM enables local medical personnel with limited specialist training to perform advanced procedures with real-time intelligent guidance.

Approach

My contribution focuses on the lower-limb DVT ultrasound scan workflow. The system delivers just-in-time multimodal interventions — visual, auditory, and haptic cues — to guide practitioners through each step of the scan without requiring a remote specialist.

A core component is a real-time fine-grained AI assistant that instructs the practitioner on precise probe movements to capture optimal ultrasound images. This involves:

• Data collection & annotation — Capturing expert scan sessions and labeling fine-grained probe positions and movements to build a training dataset from the ground up
• Foundational model training — Training perception models that can interpret live ultrasound imagery and map it to actionable probe movement instructions
• System integration — Developing and maintaining the guidance pipeline using ROS2 for real-time sensor communication, Intel RealSense for depth sensing, and PyTorch for model inference

Role

I am responsible for developing, maintaining, and integrating the different guidance components of the DVT scan workflow — spanning perception, decision-making, and real-time instruction delivery — in close collaboration with clinical partners and the broader PARADIGM team.

Impact

This work extends the human-AI teaming paradigm from aerospace into healthcare, applying lessons from real-time AI assistance for disorienting control tasks to a setting where errors carry direct patient safety implications. The goal is an assistant that not only improves task outcomes but builds practitioner competence over time.