T2ID-CAS: Diffusion Model and Class Aware Sampling to Mitigate Class Imbalance in Neck Ultrasound Anatomical Landmark...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ultrasound Technology

Background:

• Neck ultrasound (US) is crucial for real-time airway management.
• Deep learning models for anatomical landmark detection in neck US face challenges due to class imbalance, particularly for underrepresented structures like tracheal rings and vocal folds.
• This imbalance hinders the efficiency and accuracy of AI-assisted interventions.

Purpose of the Study:

• To develop and evaluate a hybrid approach (T2ID-CAS) to mitigate class imbalance in neck US datasets.
• To improve the performance of deep learning models for anatomical landmark detection in neck US.
• To enhance the reliability of AI-assisted ultrasound-guided airway management.

Main Methods:

• Proposed T2ID-CAS, a hybrid method combining a text-to-image latent diffusion model with class-aware sampling.
• Generated high-quality synthetic data for underrepresented classes in neck US datasets.
• Utilized YOLOv9 for anatomical landmark detection on neck US images.

Main Results:

• T2ID-CAS achieved a mean Average Precision (mAP) of 88.2 for anatomical landmark detection.
• This significantly outperformed the baseline method, which achieved an mAP of 66.
• The approach demonstrated effectiveness in improving the representation of minority classes.

Conclusions:

• T2ID-CAS is a computationally efficient and scalable solution for addressing class imbalance in AI-assisted ultrasound.
• The method enhances the accurate detection of critical airway structures, improving ultrasound-guided airway management.
• This leads to increased safety, reduced misplacement risks, and supports precise real-time assessment in critical care settings.