Personalised Probe Guidance via Reinforcement Learning in Low-Resource Settings
By Saraaz Khalil - Radguide LTD
Objective:
To develop and evaluate a reinforcement learning (RL)-based system that provides real-time, personalised probe guidance for non-specialist ultrasound operators in low-resource environments, with the aim of improving scan quality and diagnostic reliability.
Introduction:
Ultrasound access is expanding globally, but image acquisition remains operator-dependent, particularly in rural or underserved regions where trained sonographers are scarce. Reinforcement learning presents an opportunity to autonomously assist operators in achieving optimal probe positioning, especially where training infrastructure is limited.
Methods:
A Deep Q-Network (DQN) reinforcement learning model was trained using a dataset of 20,000 expertannotated abdominal ultrasound scans, covering common obstetric and abdominal windows. The model learned to predict real-time probe movement recommendations (e.g., tilt, rotate, shift) that
maximise image quality.
This algorithm was integrated into a low-cost, battery-operated handheld ultrasound system with embedded haptic feedback — specifically, directional vibrations to guide user adjustments. Eighteen rural health workers in Uganda, with no formal sonography training, participated in a prospective field
trial. Each operator performed standardised abdominal scans with and without RL guidance.
Key metrics included:
• Scan quality score, rated blindly by three expert radiologists on a 5-point Likert scale
• Time to image acquisition
• Diagnostic concordance with gold-standard images
Results:
RL guidance improved mean scan quality from 2.1/5 to 4.3/5 (p < 0.01). Average acquisition time decreased from 74 seconds to 47 seconds — a 37% reduction. Diagnostic concordance with expertreviewed images rose from 62% to 89%, indicating significant gains in clinical reliability.
Conclusion:
Reinforcement learning-based probe guidance empowers non-specialists to perform high-quality ultrasound, offering a scalable solution for improving diagnostic equity in low-resource and remote healthcare settings.