Efficient Ultrasound Image Analysis Models with Sonographer Gaze Assisted Distillation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Automated medical image analysis often uses large, resource-intensive deep learning models.
• Reducing model size is critical for efficient, real-time decision-making in medical imaging.
• Current methods primarily rely on manual annotations for expert knowledge integration.

Purpose of the Study:

• To develop memory-efficient deep learning models for ultrasound image analysis.
• To integrate expert sonographer's point-of-gaze data into model training.
• To improve the efficiency of automated ultrasound image analysis without compromising performance.

Main Methods:

• Utilized teacher-student knowledge transfer models.
• Combined ultrasound video data with sonographer point-of-gaze tracking.
• Focused on frame classification for fetal anatomy (abdomen, head, femur).

Main Results:

• Developed memory-efficient models for ultrasound image analysis.
• Achieved performance within 5% of conventional models.
• Models were approximately 1000 times smaller than traditional counterparts.

Conclusions:

• Expert gaze data can effectively train memory-efficient deep learning models.
• This approach significantly reduces computational and memory requirements for ultrasound analysis.
• Enables more efficient automated decision-making in medical imaging applications.