Domain-aware contrastive learning for ultrasound hip image analysis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Early diagnosis of Developmental Dysplasia of Hip (DDH) is crucial for effective treatment.
• Handheld ultrasound probes offer portability and affordability for DDH screening.
• Lower image quality from handheld probes can hinder accurate diagnosis and AI interpretation.

Purpose of the Study:

• To develop an image translation technique to enhance ultrasound images from handheld probes.
• To improve the generalization of AI-based interpretation across different ultrasound probes.
• To create pseudo-3D ultrasound images (3DUS) mimicking conventional probe quality.

Main Methods:

• Developed a novel domain-aware contrastive unpaired translation (D-CUT) technique.
• Integrated a Bone Probability Map (BPM) into the loss function to enhance bony region similarity.
• Translated 575 handheld probe images to generate pseudo-3DUS images.

Main Results:

• Generated pseudo-3DUS images comparable to conventional probes (Fréchet Inception Distance = 92).
• Achieved high structural similarity (SSIM = 0.68, Cosine Similarity = 0.65) between translated and original images.
• Demonstrated improved contrast around bony regions in translated images.

Conclusions:

• The D-CUT technique is feasible for improving ultrasound image quality from handheld probes.
• Enhanced images can support wider adoption of ultrasound for DDH screening.
• This approach could generalize AI performance across various ultrasound devices.