Autonomous Ultrasound AI for Extreme Environments

Abstract

Internal hemorrhaging remains a leading cause of preventable death in trauma scenarios, yet the Focused Assessment with Sonography in Trauma (FAST) protocol requires specialized sonographic expertise that is often unavailable in remote or extreme environments. Non-experts often struggle with the complex landmark identification and proper probe orientation, leading to diagnostic delays and catastrophic patient outcomes in isolated settings where rapid intervention is critical to patient survival.

Celestia AI is a platform-agnostic Software as a Medical Device (SaMD) that provides real-time, prescriptive navigational guidance for non-expert ultrasound operators. The system utilizes a deep UNET Convolutional Neural Network (CNN) for Right Upper Quadrant (RUQ) landmark identification, optimized through a sophisticated hybrid loss function of DICE loss and weighted cross entropy to ensure anatomical sensitivity. The software is optimized for mobile deployment, ensuring low-latency processing and responsive HMI feedback on commercial-off-the-shelf tablets.

Technical validation demonstrated an 8-bit inference pipeline achieving ≤ 3s latency while maintaining a diagnostic sensitivity of 95% ± 5% against board-certified clinical datasets. Design verification confirmed the system operates within a 40°C thermal safety limit during continuous processing to ensure hardware longevity and patient safety. Usability testing indicates that prescriptive UI cues significantly reduce the cognitive load and orientation errors for novice users compared to traditional ultrasound interfaces.

Celestia AI bridges the specialized expertise gap, empowering generalists to perform expert-level trauma assessments in disaster relief, rural medicine, and extraterrestrial missions. By providing a validated, low-latency diagnostic aid, Celestia improves global health equity and patient survivability in high-stakes, resource-limited environments.

Video