Active Inference and Deep Generative Modeling for Cognitive Ultrasound.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Ultrasound (US) offers accessible medical imaging but suffers from operator and patient-dependent image quality, limiting diagnostic efficacy in difficult cases.
• Current US systems lack autonomous adaptation, leading to suboptimal image acquisition and diagnostic uncertainty.

Purpose of the Study:

• To reframe ultrasound systems as information-seeking agents capable of autonomous, personalized imaging.
• To leverage deep generative models and Bayesian inference for adaptive ultrasound acquisition.
• To enhance diagnostic value and image quality, particularly in challenging patient scenarios.

Main Methods:

• Interpreting ultrasound data acquisition as a perception-action loop where systems learn from interactions with the anatomical environment.
• Employing Bayesian inference with deep generative models to jointly optimize action (data acquisition) and perception (state inference).
• Developing cognitive, closed-loop ultrasound systems that utilize adaptive beamsteering and scanline selection based on learned anatomical models.

Main Results:

• Demonstrated ultrasound systems acting as information-seeking agents that personalize imaging sequences.
• Showcased adaptive beamsteering and scanline selection driven by deep generative models.
• Illustrated the potential for maximizing information gain and reducing uncertainty in ultrasound imaging.

Conclusions:

• Recasting ultrasound systems as cognitive agents with generative AI can significantly improve image quality and diagnostic accuracy.
• Autonomous, adaptive imaging strategies enhance information gain and reduce operator/patient dependency.
• This approach holds promise for overcoming limitations in difficult-to-image patients and advancing the field of medical ultrasound.