An orchestration learning framework for ultrasound imaging: Prompt-Guided Hyper-Perception and Attention-Matching...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Ultrasound imaging is crucial for diagnostics but manual interpretation is time-consuming and variable.
• Single-task AI and current foundation models face challenges with diverse medical ultrasound data.
• Limitations include noise, data variability, and difficulty adapting prior knowledge to specific tasks.

Purpose of the Study:

• To introduce PerceptGuide, an orchestration learning framework for general-purpose ultrasound classification and segmentation.
• To address limitations of existing AI solutions in handling real-world medical ultrasound datasets.
• To develop a versatile AI framework adaptable to diverse ultrasound imaging tasks and organs.

Main Methods:

• Developed PerceptGuide, an orchestration learning framework utilizing prompted hyper-perception.
• Employed supervised pre-training for direct capture of task-relevant features, avoiding extensive fine-tuning.
• Utilized a novel orchestration mechanism with Object, Task, Input, and Position prompts for adaptability.
• Introduced a downstream synchronization training stage for enhanced generalization.
• Compiled a large-scale Multi-task, Multi-organ public ultrasound dataset (M²-US) for research.

Main Results:

• PerceptGuide demonstrated robustness and versatility in multi-task, multi-organ ultrasound image processing.
• Outperformed specialist AI models and existing general AI solutions.
• Improved segmentation accuracy from 82.26% to 86.45%.
• Enhanced classification accuracy from 71.30% to 79.08%.
• Significantly reduced model parameters compared to specialist models.

Conclusions:

• PerceptGuide offers a superior, general-purpose solution for ultrasound image analysis.
• The framework's prompted hyper-perception mechanism effectively adapts to diverse ultrasound data and tasks.
• Achieved significant improvements in both classification and segmentation, with reduced computational cost.
• Shows strong potential for real-world clinical applications in medical imaging.