Neural Network Model of eFAST Target Prediction for Robotic Ultrasound Diagnostics in Austere Environments

Abstract

Modern robotic technology has the potential to solve complex problems in healthcare, such as providing technology to support medical care in austere environments characterized by the restricted availability of local medical professionals or with high patient-to-caregiver ratios. Medical robots can be deployed as a force multiplier in situations when skilled healthcare providers are limited and can reduce the risk of medical failures during diagnostic and intervention procedures. One example is detecting free-flowing blood in the abdomen and pneumothorax by performing the extended Focused Assessment with Sonography for Trauma (eFAST) ultrasound diagnostics examination. We developed a semi-autonomous robotic ultrasound system that intelligently perceives the patient’s pose and determines the corresponding eFAST configuration for the robotic-assisted procedure. The dynamic pose of the patient is identified in real-time using an optimization-based algorithm, while the eFAST configurations are predicted by a neural network model. A robot manipulator holding an ultrasound device is autonomously driven to the body-normal vector of each eFAST target. This approach accomplishes robust prediction for non-linear problems, even with dynamic posture of the detected body in highly unstructured sites with variable patient shape and pose.